com.facebook.hive.orc.WriterImpl.java Source code

Java tutorial

Introduction

Here is the source code for com.facebook.hive.orc.WriterImpl.java

Source

//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.

/**
 * Licensed to the Apache Software Foundation (ASF) under one
 * or more contributor license agreements.  See the NOTICE file
 * distributed with this work for additional information
 * regarding copyright ownership.  The ASF licenses this file
 * to you under the Apache License, Version 2.0 (the
 * "License"); you may not use this file except in compliance
 * with the License.  You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

package com.facebook.hive.orc;

import java.io.IOException;
import java.io.OutputStream;
import java.nio.ByteBuffer;
import java.sql.Timestamp;
import java.util.ArrayList;
import java.util.HashMap;
import java.util.HashSet;
import java.util.List;
import java.util.Map;
import java.util.Random;
import java.util.Set;
import java.util.TreeMap;

import com.facebook.hive.orc.compression.CompressionCodec;
import com.facebook.hive.orc.compression.CompressionKind;
import com.facebook.hive.orc.compression.SnappyCodec;
import com.facebook.hive.orc.compression.ZlibCodec;
import com.facebook.hive.orc.statistics.ColumnStatisticsImpl;
import org.apache.commons.logging.Log;
import org.apache.commons.logging.LogFactory;
import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.fs.FSDataOutputStream;
import org.apache.hadoop.fs.FileSystem;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.hive.common.io.RawDatasizeConst;
import org.apache.hadoop.hive.serde2.ReaderWriterProfiler;
import org.apache.hadoop.hive.serde2.objectinspector.ListObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.MapObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.ObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.PrimitiveObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.StructField;
import org.apache.hadoop.hive.serde2.objectinspector.StructObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.UnionObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.primitive.BinaryObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.primitive.BooleanObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.primitive.ByteObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.primitive.DoubleObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.primitive.FloatObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.primitive.IntObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.primitive.LongObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.primitive.ShortObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.primitive.StringObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.primitive.TimestampObjectInspector;
import org.apache.hadoop.io.BytesWritable;
import org.apache.hadoop.io.Text;

import com.facebook.hive.orc.OrcConf.ConfVars;
import com.facebook.hive.orc.OrcProto.Stream.Kind;
import com.google.protobuf.ByteString;
import com.google.protobuf.CodedOutputStream;

/**
 * An ORC file writer. The file is divided into stripes, which is the natural
 * unit of work when reading. Each stripe is buffered in memory until the
 * memory reaches the stripe size and then it is written out broken down by
 * columns. Each column is written by a TreeWriter that is specific to that
 * type of column. TreeWriters may have children TreeWriters that handle the
 * sub-types. Each of the TreeWriters writes the column's data as a set of
 * streams.
 *
 * This class is synchronized so that multi-threaded access is ok. In
 * particular, because the MemoryManager is shared between writers, this class
 * assumes that checkMemory may be called from a separate thread.
 */
public class WriterImpl implements Writer, MemoryManager.Callback {

    private static final Log LOG = LogFactory.getLog(WriterImpl.class);

    private static final int HDFS_BUFFER_SIZE = 256 * 1024;
    private static final int MIN_ROW_INDEX_STRIDE = 1000;

    public static final int SHORT_BYTE_SIZE = 2;
    public static final int INT_BYTE_SIZE = 4;
    public static final int LONG_BYTE_SIZE = 8;

    // Specifies how many index entries are created for a present stream
    public static final int UNCOMPRESSED_PRESENT_STREAM_INDEX_ENTRIES = 3;
    public static final int COMPRESSED_PRESENT_STREAM_INDEX_ENTRIES = 4;

    private final FileSystem fs;
    private final Path path;
    private final long stripeSize;
    private final int rowIndexStride;
    private final CompressionKind compress;
    private final CompressionCodec codec;
    private final int bufferSize;
    // the streams that make up the current stripe
    private final Map<StreamName, BufferedStream> streams = new TreeMap<StreamName, BufferedStream>();

    private FSDataOutputStream rawWriter = null;
    // the compressed metadata information outStream
    private OutStream writer = null;
    // a protobuf outStream around streamFactory
    private CodedOutputStream protobufWriter = null;
    private long headerLength;
    private int columnCount;
    private long rowCount = 0;
    private long rowsInStripe = 0;
    private int rowsInIndex = 0;
    private long rawDataSize = 0;
    private final List<OrcProto.StripeInformation> stripes = new ArrayList<OrcProto.StripeInformation>();
    private final Map<String, ByteString> userMetadata = new TreeMap<String, ByteString>();
    private final StreamFactory streamFactory = new StreamFactory();
    private final TreeWriter treeWriter;
    private final OrcProto.RowIndex.Builder rowIndex = OrcProto.RowIndex.newBuilder();
    private final boolean buildIndex;
    private final MemoryManager memoryManager;
    private final boolean useVInts;
    private final int dfsBytesPerChecksum;
    private final long maxDictSize;
    private final MemoryEstimate memoryEstimate = new MemoryEstimate();

    private final Configuration conf;

    WriterImpl(FileSystem fs, Path path, Configuration conf, ObjectInspector inspector, long stripeSize,
            CompressionKind compress, int bufferSize, int rowIndexStride, MemoryManager memoryManager)
            throws IOException {
        this.fs = fs;
        this.path = path;
        this.conf = conf;
        this.stripeSize = stripeSize;
        this.compress = compress;
        this.bufferSize = bufferSize;
        this.rowIndexStride = rowIndexStride;
        this.memoryManager = memoryManager;
        buildIndex = rowIndexStride > 0;
        codec = createCodec(compress, conf);
        useVInts = OrcConf.getBoolVar(conf, OrcConf.ConfVars.HIVE_ORC_USE_VINTS);
        treeWriter = createTreeWriter(inspector, streamFactory, false, conf, useVInts,
                memoryManager.isLowMemoryMode(), memoryEstimate);
        dfsBytesPerChecksum = conf.getInt("io.bytes.per.checksum", 512);
        if (buildIndex && rowIndexStride < MIN_ROW_INDEX_STRIDE) {
            throw new IllegalArgumentException("Row stride must be at least " + MIN_ROW_INDEX_STRIDE);
        }
        maxDictSize = OrcConf.getLongVar(conf, OrcConf.ConfVars.HIVE_ORC_MAX_DICTIONARY_SIZE);
        // ensure that we are able to handle callbacks before we register ourselves
        memoryManager.addWriter(path, stripeSize, this, memoryEstimate.getTotalMemory());
    }

    static CompressionCodec createCodec(CompressionKind kind) {
        // To be used for cases where we don't care about configuring the codec,
        // e.g. reads
        return createCodec(kind, null);
    }

    static CompressionCodec createCodec(CompressionKind kind, Configuration conf) {
        switch (kind) {
        case NONE:
            return null;
        case ZLIB:
            return new ZlibCodec(conf);
        case SNAPPY:
            return new SnappyCodec();
        case LZO:
            try {
                Class<? extends CompressionCodec> lzo = (Class<? extends CompressionCodec>) Class
                        .forName("com.facebook.hive.orc.LzoCodec");
                return lzo.newInstance();
            } catch (ClassNotFoundException e) {
                throw new IllegalArgumentException("LZO is not available.", e);
            } catch (InstantiationException e) {
                throw new IllegalArgumentException("Problem initializing LZO", e);
            } catch (IllegalAccessException e) {
                throw new IllegalArgumentException("Insufficient access to LZO", e);
            }
        default:
            throw new IllegalArgumentException("Unknown compression codec: " + kind);
        }
    }

    @Override
    public synchronized void enterLowMemoryMode() throws IOException {
        // Don't use dictionaries
        treeWriter.abandonDictionaries();
        // If the Zlib compression level is less than 6, raise it to 6 to compensate for the fact
        // we aren't using dictionaries
        if (codec != null && OrcConf.getIntVar(conf, ConfVars.HIVE_ORC_ZLIB_COMPRESSION_LEVEL) < 6) {
            OrcConf.setIntVar(conf, ConfVars.HIVE_ORC_ZLIB_COMPRESSION_LEVEL, 6);
            codec.reloadConfigurations(conf);
        }
    }

    /**
     * This class is used to hold the contents of streams as they are buffered.
     * The TreeWriters write to the outStream and the codec compresses the
     * data as buffers fill up and stores them in the output list. When the
     * stripe is being written, the whole stream is written to the file.
     */
    private class BufferedStream implements OutStream.OutputReceiver {
        private final OutStream outStream;
        private final List<ByteBuffer> output = new ArrayList<ByteBuffer>();

        BufferedStream(String name, int bufferSize, CompressionCodec codec) throws IOException {
            outStream = new OutStream(name, bufferSize, codec, this, memoryEstimate);
        }

        /**
         * Receive a buffer from the compression codec.
         * @param buffer the buffer to save
         * @throws IOException
         */
        @Override
        public void output(ByteBuffer buffer) {
            output.add(buffer);
        }

        /**
         * Get the number of bytes in buffers that are allocated to this stream.
         * @return number of bytes in buffers
         */
        public long getBufferSize() {
            long result = 0;
            for (ByteBuffer buf : output) {
                result += buf.capacity();
            }
            return outStream.getBufferSize() + result;
        }

        /**
         * Flush the stream to the codec.
         * @throws IOException
         */
        public void flush(boolean reuseBuffer) throws IOException {
            outStream.flush(reuseBuffer);
        }

        /**
         * Clear all of the buffers.
         * @throws IOException
         */
        public void clear() throws IOException {
            outStream.clear();
            for (ByteBuffer buf : output) {
                memoryEstimate.decrementTotalMemory(buf.capacity());
            }
            output.clear();
        }

        /**
         * Check the state of suppress flag in output stream
         * @return value of suppress flag
         */
        public boolean isSuppressed() {
            return outStream.isSuppressed();
        }

        /**
         * Write the saved compressed buffers to the OutputStream.
         * @param out the stream to write to
         * @throws IOException
         */
        void spillTo(OutputStream out) throws IOException {
            for (ByteBuffer buffer : output) {
                out.write(buffer.array(), buffer.arrayOffset() + buffer.position(), buffer.remaining());
            }
        }

        @Override
        public String toString() {
            return outStream.toString();
        }
    }

    /**
     * An output receiver that writes the ByteBuffers to the output stream
     * as they are received.
     */
    private class DirectStream implements OutStream.OutputReceiver {
        private final FSDataOutputStream output;

        DirectStream(FSDataOutputStream output) {
            this.output = output;
        }

        @Override
        public void output(ByteBuffer buffer) throws IOException {
            output.write(buffer.array(), buffer.arrayOffset() + buffer.position(), buffer.remaining());
            memoryEstimate.decrementTotalMemory(buffer.capacity());
        }
    }

    static class RowIndexPositionRecorder implements PositionRecorder {
        private final OrcProto.RowIndexEntry.Builder builder;

        RowIndexPositionRecorder(OrcProto.RowIndexEntry.Builder builder) {
            this.builder = builder;
        }

        @Override
        public void addPosition(long position) {
            builder.addPositions(position);
        }
    }

    /**
     * Interface from the Writer to the TreeWriters. This limits the visibility
     * that the TreeWriters have into the Writer.
     */
    private class StreamFactory {
        /**
         * Create a stream to store part of a column.
         * @param column the column id for the stream
         * @param kind the kind of stream
         * @return The output outStream that the section needs to be written to.
         * @throws IOException
         */
        public OutStream createStream(int column, OrcProto.Stream.Kind kind) throws IOException {
            StreamName name = new StreamName(column, kind);
            BufferedStream result = streams.get(name);
            if (result == null) {
                result = new BufferedStream(name.toString(), bufferSize, codec);
                streams.put(name, result);
            }
            return result.outStream;
        }

        /**
         * Get the next column id.
         * @return a number from 0 to the number of columns - 1
         */
        public int getNextColumnId() {
            return columnCount++;
        }

        /**
         * Get the stride rate of the row index.
         */
        public int getRowIndexStride() {
            return rowIndexStride;
        }

        /**
         * Should be building the row index.
         * @return true if we are building the index
         */
        public boolean buildIndex() {
            return buildIndex;
        }

        /**
         * Is the ORC file compressed?
         * @return are the streams compressed
         */
        public boolean isCompressed() {
            return codec != null;
        }
    }

    /**
     * The parent class of all of the writers for each column. Each column
     * is written by an instance of this class. The compound types (struct,
     * list, map, and union) have children tree writers that write the children
     * types.
     */
    private abstract static class TreeWriter {
        protected final int id;
        protected final ObjectInspector inspector;
        private final BitFieldWriter isPresent;
        private final boolean isCompressed;
        protected final ColumnStatisticsImpl indexStatistics;
        private final ColumnStatisticsImpl fileStatistics;
        protected TreeWriter[] childrenWriters;
        protected final RowIndexPositionRecorder rowIndexPosition;
        private final OrcProto.RowIndex.Builder rowIndex;
        private final OrcProto.RowIndexEntry.Builder rowIndexEntry;
        private final PositionedOutputStream rowIndexStream;
        private final Configuration conf;
        protected long stripeRawDataSize = 0;
        protected long rowRawDataSize = 0;
        protected final boolean useVInts;
        private int numStripes = 0;
        private boolean foundNulls;
        private OutStream isPresentOutStream;
        protected final MemoryEstimate memoryEstimate;

        /**
         * Create a tree writer
         * @param columnId the column id of the column to write
         * @param inspector the object inspector to use
         * @param streamFactory limited access to the Writer's data.
         * @param nullable can the value be null?
         * @throws IOException
         */
        TreeWriter(int columnId, ObjectInspector inspector, StreamFactory streamFactory, boolean nullable,
                Configuration conf, boolean useVInts, MemoryEstimate memoryEstimate) throws IOException {
            this.id = columnId;
            this.inspector = inspector;
            this.conf = conf;
            this.useVInts = useVInts;
            this.memoryEstimate = memoryEstimate;
            this.isCompressed = streamFactory.isCompressed();
            if (nullable) {
                isPresentOutStream = streamFactory.createStream(id, OrcProto.Stream.Kind.PRESENT);
                isPresent = new BitFieldWriter(isPresentOutStream, 1);
            } else {
                isPresent = null;
            }
            this.foundNulls = false;
            indexStatistics = ColumnStatisticsImpl.create(inspector);
            fileStatistics = ColumnStatisticsImpl.create(inspector);
            childrenWriters = new TreeWriter[0];
            rowIndex = OrcProto.RowIndex.newBuilder();
            rowIndexEntry = OrcProto.RowIndexEntry.newBuilder();
            rowIndexPosition = new RowIndexPositionRecorder(rowIndexEntry);
            if (streamFactory.buildIndex()) {
                rowIndexStream = streamFactory.createStream(id, OrcProto.Stream.Kind.ROW_INDEX);
            } else {
                rowIndexStream = null;
            }
        }

        protected int getNumStripes() {
            return numStripes;
        }

        protected OrcProto.RowIndex.Builder getRowIndex() {
            return rowIndex;
        }

        protected ColumnStatisticsImpl getFileStatistics() {
            return fileStatistics;
        }

        protected OrcProto.RowIndexEntry.Builder getRowIndexEntry() {
            return rowIndexEntry;
        }

        /**
         * Add a new value to the column.
         * @param obj
         * @throws IOException
         */
        abstract void write(Object obj) throws IOException;

        void write(Object obj, long rawDataSize) throws IOException {

            if (obj != null) {
                setRawDataSize(rawDataSize);
            } else {
                // Estimate the raw size of null as 1 byte
                setRawDataSize(RawDatasizeConst.NULL_SIZE);
            }

            flushRow(obj);
        }

        /**
         * Update the row count and mark the isPresent bit
         */
        void flushRow(Object obj) throws IOException {
            if (obj != null) {
                indexStatistics.increment();
            }

            if (isPresent != null) {
                isPresent.write(obj == null ? 0 : 1);
                if (obj == null) {
                    foundNulls = true;
                }
            }
        }

        private void removeIsPresentPositions() {
            for (int i = 0; i < rowIndex.getEntryCount(); ++i) {
                OrcProto.RowIndexEntry.Builder entry = rowIndex.getEntryBuilder(i);
                List<Long> positions = entry.getPositionsList();
                // bit streams use 3 positions if uncompressed, 4 if compressed
                positions = positions.subList(isCompressed ? COMPRESSED_PRESENT_STREAM_INDEX_ENTRIES
                        : UNCOMPRESSED_PRESENT_STREAM_INDEX_ENTRIES, positions.size());
                entry.clearPositions();
                entry.addAllPositions(positions);
            }
        }

        /**
         * Sets the row raw data size and updates the stripe raw data size
         *
         * @param rawDataSize
         */
        protected void setRawDataSize(long rawDataSize) {
            rowRawDataSize = rawDataSize;
            stripeRawDataSize += rawDataSize;
        }

        /**
         * Write the stripe out to the file.
         * @param builder the stripe footer that contains the information about the
         *                layout of the stripe. The TreeWriter is required to update
         *                the footer with its information.
         * @param requiredIndexEntries the number of index entries that are
         *                             required. this is to check to make sure the
         *                             row index is well formed.
         * @throws IOException
         */
        void writeStripe(OrcProto.StripeFooter.Builder builder, int requiredIndexEntries) throws IOException {
            numStripes++;
            if (isPresent != null) {
                isPresent.flush();

                // if no nulls are found in a stream, then suppress the stream
                if (!foundNulls) {
                    isPresentOutStream.suppress();
                    // since isPresent bitstream is suppressed, update the index to
                    // remove the positions of the isPresent stream
                    if (rowIndexStream != null) {
                        removeIsPresentPositions();
                    }
                }
            }

            // reset the flag for next stripe
            foundNulls = false;

            builder.addColumns(getEncoding());
            if (rowIndexStream != null) {
                if (rowIndex.getEntryCount() != requiredIndexEntries) {
                    throw new IllegalArgumentException("Column has wrong number of " + "index entries found: "
                            + rowIndex.getEntryCount() + " expected: " + requiredIndexEntries);
                }
                rowIndex.build().writeTo(rowIndexStream);
                rowIndexStream.flush();
            }
            rowIndex.clear();
            rowIndexEntry.clear();
            stripeRawDataSize = 0;
        }

        TreeWriter[] getChildrenWriters() {
            return childrenWriters;
        }

        /**
         * For all the TreeWriters that buffer rows, process
         * all the buffered rows.
         */
        void flush() throws IOException {
            for (TreeWriter writer : childrenWriters) {
                writer.flush();
            }
            return;
        }

        /**
         * Get the encoding for this column.
         * @return the information about the encoding of this column
         */
        OrcProto.ColumnEncoding getEncoding() {
            return OrcProto.ColumnEncoding.newBuilder().setKind(OrcProto.ColumnEncoding.Kind.DIRECT).build();
        }

        /**
         * Create a row index entry with the previous location and the current
         * index statistics. Also merges the index statistics into the file
         * statistics before they are cleared. Finally, it records the start of the
         * next index and ensures all of the children columns also create an entry.
         * @throws IOException
         */
        void createRowIndexEntry() throws IOException {
            fileStatistics.merge(indexStatistics);
            rowIndexEntry.setStatistics(indexStatistics.serialize());
            indexStatistics.reset();
            rowIndex.addEntry(rowIndexEntry);
            rowIndexEntry.clear();
            recordPosition(rowIndexPosition);
            for (TreeWriter child : childrenWriters) {
                child.createRowIndexEntry();
            }
        }

        /**
         * Record the current position in each of this column's streams.
         * @param recorder where should the locations be recorded
         * @throws IOException
         */
        void recordPosition(PositionRecorder recorder) throws IOException {
            if (isPresent != null) {
                isPresent.getPosition(recorder);
            }
        }

        public void abandonDictionaries() throws IOException {
            for (TreeWriter child : childrenWriters) {
                child.abandonDictionaries();
            }
        }

        long getStripeRawDataSize() {
            return stripeRawDataSize;
        }

        long getRowRawDataSize() {
            return rowRawDataSize;
        }
    }

    private static class BooleanTreeWriter extends TreeWriter {
        private final BitFieldWriter writer;

        BooleanTreeWriter(int columnId, ObjectInspector inspector, StreamFactory writer, boolean nullable,
                Configuration conf, boolean useVInts, boolean lowMemoryMode, MemoryEstimate memoryEstimate)
                throws IOException {
            super(columnId, inspector, writer, nullable, conf, useVInts, memoryEstimate);
            PositionedOutputStream out = writer.createStream(id, OrcProto.Stream.Kind.DATA);
            this.writer = new BitFieldWriter(out, 1);
            recordPosition(rowIndexPosition);
        }

        @Override
        void write(Object obj) throws IOException {
            super.write(obj, RawDatasizeConst.BOOLEAN_SIZE);
            if (obj != null) {
                boolean val = ((BooleanObjectInspector) inspector).get(obj);
                indexStatistics.updateBoolean(val);
                writer.write(val ? 1 : 0);
            }
        }

        @Override
        void writeStripe(OrcProto.StripeFooter.Builder builder, int requiredIndexEntries) throws IOException {
            super.writeStripe(builder, requiredIndexEntries);
            writer.flush();
            recordPosition(rowIndexPosition);
        }

        @Override
        void recordPosition(PositionRecorder recorder) throws IOException {
            super.recordPosition(recorder);
            writer.getPosition(recorder);
        }
    }

    private static class ByteTreeWriter extends TreeWriter {
        private final RunLengthByteWriter writer;

        ByteTreeWriter(int columnId, ObjectInspector inspector, StreamFactory writer, boolean nullable,
                Configuration conf, boolean useVInts, boolean lowMemoryMode, MemoryEstimate memoryEstimate)
                throws IOException {
            super(columnId, inspector, writer, nullable, conf, useVInts, memoryEstimate);
            this.writer = new RunLengthByteWriter(writer.createStream(id, OrcProto.Stream.Kind.DATA));
            recordPosition(rowIndexPosition);
        }

        @Override
        void write(Object obj) throws IOException {
            super.write(obj, RawDatasizeConst.BYTE_SIZE);
            if (obj != null) {
                byte val = ((ByteObjectInspector) inspector).get(obj);
                indexStatistics.updateInteger(val);
                writer.write(val);
            }
        }

        @Override
        void writeStripe(OrcProto.StripeFooter.Builder builder, int requiredIndexEntries) throws IOException {
            super.writeStripe(builder, requiredIndexEntries);
            writer.flush();
            recordPosition(rowIndexPosition);
        }

        @Override
        void recordPosition(PositionRecorder recorder) throws IOException {
            super.recordPosition(recorder);
            writer.getPosition(recorder);
        }
    }

    private static class IntegerTreeWriter extends TreeWriter {
        private final PositionedOutputStream output;
        private DynamicIntArray rows;
        private final PositionedOutputStream inDictionaryStream;
        private final BitFieldWriter inDictionary;
        private final List<OrcProto.RowIndexEntry> savedRowIndex = new ArrayList<OrcProto.RowIndexEntry>();
        private final boolean buildIndex;
        private final List<Long> rowIndexValueCount = new ArrayList<Long>();
        private final float dictionaryKeySizeThreshold;

        private IntDictionaryEncoder dictionary;
        private boolean useDictionaryEncoding = true;
        private final StreamFactory writer;
        private final int numBytes;
        private final Long[] buffer;
        private int bufferIndex = 0;
        private long bufferedBytes = 0;
        private final int recomputeStripeEncodingInterval;
        PositionedOutputStream rowOutput;
        private boolean abandonDictionaries = false;
        private final boolean sortKeys;
        private int dictionarySize;
        private final boolean useStrideDictionaries;

        IntegerTreeWriter(int columnId, ObjectInspector inspector, StreamFactory writerFactory, boolean nullable,
                Configuration conf, boolean useVInts, int numBytes, boolean lowMemoryMode,
                MemoryEstimate memoryEstimate) throws IOException {
            super(columnId, inspector, writerFactory, nullable, conf, useVInts, memoryEstimate);
            writer = writerFactory;
            sortKeys = OrcConf.getBoolVar(conf, OrcConf.ConfVars.HIVE_ORC_DICTIONARY_SORT_KEYS);
            useStrideDictionaries = OrcConf.getBoolVar(conf, OrcConf.ConfVars.HIVE_ORC_BUILD_STRIDE_DICTIONARY);

            this.numBytes = numBytes;
            recomputeStripeEncodingInterval = OrcConf.getIntVar(conf,
                    OrcConf.ConfVars.HIVE_ORC_DICTIONARY_ENCODING_INTERVAL);

            if (!lowMemoryMode) {
                dictionary = new IntDictionaryEncoder(sortKeys, numBytes, useVInts, memoryEstimate);
                rows = new DynamicIntArray(memoryEstimate);
            } else {
                abandonDictionaries = true;
                rowOutput = writer.createStream(id, OrcProto.Stream.Kind.DATA);
                useDictionaryEncoding = false;
            }
            output = writer.createStream(id, OrcProto.Stream.Kind.DICTIONARY_DATA);
            inDictionaryStream = writer.createStream(id, OrcProto.Stream.Kind.IN_DICTIONARY);
            inDictionary = new BitFieldWriter(inDictionaryStream, 1);

            dictionaryKeySizeThreshold = OrcConf.getFloatVar(conf,
                    OrcConf.ConfVars.HIVE_ORC_DICTIONARY_NUMERIC_KEY_SIZE_THRESHOLD);

            int bufferLength = OrcConf.getIntVar(conf, OrcConf.ConfVars.HIVE_ORC_ROW_BUFFER_SIZE);

            buffer = new Long[bufferLength];

            recordPosition(rowIndexPosition);
            rowIndexValueCount.add(0L);
            buildIndex = writer.buildIndex();
            if (buildIndex && lowMemoryMode) {
                rowOutput.getPosition(rowIndexPosition);
            }
        }

        boolean determineEncodingStripe() {
            return (getNumStripes() % recomputeStripeEncodingInterval) == 0 && !abandonDictionaries;
        }

        @Override
        void write(Object obj) throws IOException {
            if (obj != null) {
                switch (inspector.getCategory()) {
                case PRIMITIVE:
                    switch (((PrimitiveObjectInspector) inspector).getPrimitiveCategory()) {
                    case SHORT:
                        buffer[bufferIndex++] = new Long(((ShortObjectInspector) inspector).get(obj));
                        setRawDataSize(RawDatasizeConst.SHORT_SIZE);
                        break;
                    case INT:
                        buffer[bufferIndex++] = new Long(((IntObjectInspector) inspector).get(obj));
                        setRawDataSize(RawDatasizeConst.INT_SIZE);
                        break;
                    case LONG:
                        buffer[bufferIndex++] = new Long(((LongObjectInspector) inspector).get(obj));
                        setRawDataSize(RawDatasizeConst.LONG_SIZE);
                        break;
                    default:
                        throw new IllegalArgumentException("Bad Category: Dictionary Encoding not available for "
                                + ((PrimitiveObjectInspector) inspector).getPrimitiveCategory());
                    }
                    break;
                default:
                    throw new IllegalArgumentException(
                            "Bad Category: DictionaryEncoding not available for " + inspector.getCategory());
                }
                // Increment the total memory for the buffered long
                memoryEstimate.incrementTotalMemory(RawDatasizeConst.LONG_SIZE);
                bufferedBytes += RawDatasizeConst.LONG_SIZE;
            } else {
                buffer[bufferIndex++] = null;
                setRawDataSize(RawDatasizeConst.NULL_SIZE);
            }
            if (bufferIndex == buffer.length) {
                flush();
            }
        }

        @Override
        void flush() throws IOException {
            for (int i = 0; i < bufferIndex; i++) {
                Long val = buffer[i];
                buffer[i] = null;
                if (val != null) {
                    if (useCarriedOverDirectEncoding()) {
                        SerializationUtils.writeIntegerType(rowOutput, val, numBytes, true, useVInts);
                    } else {
                        rows.add(dictionary.add(val));
                    }
                    indexStatistics.updateInteger(val);
                }

                super.flushRow(val);
            }
            bufferIndex = 0;
            memoryEstimate.decrementTotalMemory(bufferedBytes);
            bufferedBytes = 0;
        }

        boolean getUseDictionaryEncoding() {
            return (rows.size() > 0
                    && (float) (dictionary.size()) / (float) rows.size() <= dictionaryKeySizeThreshold);
        }

        /**
         * Returns true iff the encoding is not being determined using this stripe, and
         * the previously determined encoding was direct.
         */
        private boolean useCarriedOverDirectEncoding() {
            return !determineEncodingStripe() && !useDictionaryEncoding;
        }

        @Override
        void writeStripe(OrcProto.StripeFooter.Builder builder, int requiredIndexEntries) throws IOException {
            if (determineEncodingStripe()) {
                useDictionaryEncoding = getUseDictionaryEncoding();
            }
            if (useDictionaryEncoding) {
                rowOutput = new RunLengthIntegerWriter(writer.createStream(id, OrcProto.Stream.Kind.DATA), false,
                        INT_BYTE_SIZE, useVInts);
            } else if (determineEncodingStripe()) {
                rowOutput = writer.createStream(id, OrcProto.Stream.Kind.DATA);
            }

            final long[] dumpOrder;
            final int[] counts;
            if (useDictionaryEncoding) {
                // Traverse the dictionary keys writing out the bytes and lengths; and
                // creating the map from the original order to the final sorted order.
                dumpOrder = new long[dictionary.size()];
                counts = new int[dictionary.size()];
                dictionary.visit(new IntDictionaryEncoder.Visitor<Long>() {
                    int currentId = 0;

                    public void visit(IntDictionaryEncoder.VisitorContext<Long> context) throws IOException {
                        int count = context.getCount();
                        counts[context.getOriginalPosition()] = count;
                        if (!useStrideDictionaries || count > 1) {
                            dictionarySize++;
                            context.writeBytes(output);
                            dumpOrder[context.getOriginalPosition()] = currentId++;
                        } else {
                            dumpOrder[context.getOriginalPosition()] = dictionary
                                    .getValue(context.getOriginalPosition());
                        }
                    }
                });
            } else {
                dumpOrder = null;
                counts = null;
            }

            if (!useCarriedOverDirectEncoding()) {
                writeData(useDictionaryEncoding, dumpOrder, counts);
            }
            // we need to build the rowindex before calling super, since it
            // writes it out.
            super.writeStripe(builder, requiredIndexEntries);
            if (useDictionaryEncoding) {
                output.unsuppress();
                inDictionary.flush();
                if (dictionarySize == dictionary.size()) {
                    inDictionaryStream.suppress();
                } else {
                    inDictionaryStream.unsuppress();
                }
                output.flush();
            } else {
                output.suppress();
                inDictionaryStream.suppress();
            }
            rowOutput.flush();
            savedRowIndex.clear();
            rowIndexValueCount.clear();
            recordPosition(rowIndexPosition);
            dictionarySize = 0;
            if (useCarriedOverDirectEncoding()) {
                rowOutput = writer.createStream(id, OrcProto.Stream.Kind.DATA);
                rowOutput.getPosition(rowIndexPosition);
                if (dictionary != null) {
                    dictionary.cleanup();
                    dictionary = null;
                    rows.cleanup();
                    rows = null;
                }
            } else {
                if (dictionary == null) {
                    dictionary = new IntDictionaryEncoder(sortKeys, numBytes, useVInts, memoryEstimate);
                } else {
                    dictionary.clear();
                }
                if (rows == null) {
                    rows = new DynamicIntArray(memoryEstimate);
                } else {
                    rows.clear();
                }
            }
            rowIndexValueCount.add(0L);
        }

        private void convertDictionaryToDirect() throws IOException {
            writeData(false, null, null);
        }

        private void writeData(boolean useDictionaryEncoding, long[] dumpOrder, int[] counts) throws IOException {
            int length = rows.size();
            int rowIndexEntry = 0;
            OrcProto.RowIndex.Builder rowIndex = getRowIndex();
            // write the values translated into the dump order.
            for (int i = 0; i <= length; ++i) {
                // now that we are writing out the row values, we can finalize the
                // row index
                if (buildIndex) {
                    // If we are not using dictionary encoding and savedRowIndex is not empty, this means
                    // the conversion from dictionary encoding to direct encoding, so allow rowIndexEntry to
                    // exceed the number of saved row indeces, to create the index entry for the current
                    //stride
                    while (rowIndexEntry < rowIndexValueCount.size() && i == rowIndexValueCount.get(rowIndexEntry)
                            && (rowIndexEntry < savedRowIndex.size()
                                    || (!useDictionaryEncoding && rowIndexEntry == savedRowIndex.size()))) {
                        OrcProto.RowIndexEntry.Builder base = null;
                        RowIndexPositionRecorder recorder;
                        if (rowIndexEntry < savedRowIndex.size()) {
                            base = savedRowIndex.get(rowIndexEntry).toBuilder();
                            recorder = new RowIndexPositionRecorder(base);
                        } else {
                            recorder = rowIndexPosition;
                        }
                        if (useDictionaryEncoding && dumpOrder != null && dictionarySize != dictionary.size()) {
                            inDictionary.getPosition(recorder);
                        }
                        rowOutput.getPosition(recorder);
                        if (rowIndexEntry < savedRowIndex.size()) {
                            // If we are constructing an index entry from a saved row index, add it
                            rowIndex.addEntry(base.build());
                        }
                        rowIndexEntry++;
                    }
                }

                if (i < length) {
                    if (useDictionaryEncoding && dumpOrder != null) {
                        ((RunLengthIntegerWriter) rowOutput).write(dumpOrder[rows.get(i)]);
                        if (!useStrideDictionaries || counts[rows.get(i)] > 1) {
                            inDictionary.write(1);
                        } else {
                            inDictionary.write(0);
                        }
                    } else {
                        SerializationUtils.writeIntegerType(rowOutput, dictionary.getValue(rows.get(i)), numBytes,
                                true, useVInts);
                    }
                }
            }
        }

        @Override
        OrcProto.ColumnEncoding getEncoding() {
            if (useDictionaryEncoding) {
                return OrcProto.ColumnEncoding.newBuilder().setKind(OrcProto.ColumnEncoding.Kind.DICTIONARY)
                        .setDictionarySize(dictionarySize).build();
            } else {
                return OrcProto.ColumnEncoding.newBuilder().setKind(OrcProto.ColumnEncoding.Kind.DIRECT).build();
            }
        }

        /**
         * This method doesn't call the super method, because unlike most of the
         * other TreeWriters, this one can't record the position in the streams
         * until the stripe is being flushed. Therefore it saves all of the entries
         * and augments them with the final information as the stripe is written.
         * @throws IOException
         */
        @Override
        void createRowIndexEntry() throws IOException {
            getFileStatistics().merge(indexStatistics);
            OrcProto.RowIndexEntry.Builder rowIndexEntry = getRowIndexEntry();
            rowIndexEntry.setStatistics(indexStatistics.serialize());
            indexStatistics.reset();
            if (useCarriedOverDirectEncoding()) {
                getRowIndex().addEntry(rowIndexEntry);
            } else {
                savedRowIndex.add(rowIndexEntry.build());
            }
            rowIndexEntry.clear();
            recordPosition(rowIndexPosition);
            if (useCarriedOverDirectEncoding()) {
                rowOutput.getPosition(rowIndexPosition);
            } else {
                rowIndexValueCount.add(Long.valueOf(rows.size()));
            }
        }

        @Override
        public void abandonDictionaries() throws IOException {
            boolean useCarriedOverDirectEncoding = useCarriedOverDirectEncoding();
            abandonDictionaries = true;
            if (!useCarriedOverDirectEncoding) {
                rowOutput = writer.createStream(id, OrcProto.Stream.Kind.DATA);
                useDictionaryEncoding = false;
                convertDictionaryToDirect();
                if (rows.size() == 0) {
                    rowOutput.getPosition(rowIndexPosition);
                }
            }
            if (dictionary != null) {
                dictionary.cleanup();
                dictionary = null;
                rows.cleanup();
                rows = null;
            }
            savedRowIndex.clear();
        }
    }

    private static class FloatTreeWriter extends TreeWriter {
        private final PositionedOutputStream stream;

        FloatTreeWriter(int columnId, ObjectInspector inspector, StreamFactory writer, boolean nullable,
                Configuration conf, boolean useVInts, boolean lowMemoryMode, MemoryEstimate memoryEstimate)
                throws IOException {
            super(columnId, inspector, writer, nullable, conf, useVInts, memoryEstimate);
            this.stream = writer.createStream(id, OrcProto.Stream.Kind.DATA);
            recordPosition(rowIndexPosition);
        }

        @Override
        void write(Object obj) throws IOException {
            super.write(obj, RawDatasizeConst.FLOAT_SIZE);
            if (obj != null) {
                float val = ((FloatObjectInspector) inspector).get(obj);
                indexStatistics.updateDouble(val);
                SerializationUtils.writeFloat(stream, val);
            }
        }

        @Override
        void writeStripe(OrcProto.StripeFooter.Builder builder, int requiredIndexEntries) throws IOException {
            super.writeStripe(builder, requiredIndexEntries);
            stream.flush();
            recordPosition(rowIndexPosition);
        }

        @Override
        void recordPosition(PositionRecorder recorder) throws IOException {
            super.recordPosition(recorder);
            stream.getPosition(recorder);
        }
    }

    private static class DoubleTreeWriter extends TreeWriter {
        private final PositionedOutputStream stream;

        DoubleTreeWriter(int columnId, ObjectInspector inspector, StreamFactory writer, boolean nullable,
                Configuration conf, boolean useVInts, boolean lowMemoryMode, MemoryEstimate memoryEstimate)
                throws IOException {
            super(columnId, inspector, writer, nullable, conf, useVInts, memoryEstimate);
            this.stream = writer.createStream(id, OrcProto.Stream.Kind.DATA);
            recordPosition(rowIndexPosition);
        }

        @Override
        void write(Object obj) throws IOException {
            super.write(obj, RawDatasizeConst.DOUBLE_SIZE);
            if (obj != null) {
                double val = ((DoubleObjectInspector) inspector).get(obj);
                indexStatistics.updateDouble(val);
                SerializationUtils.writeDouble(stream, val);
            }
        }

        @Override
        void writeStripe(OrcProto.StripeFooter.Builder builder, int requiredIndexEntries) throws IOException {
            super.writeStripe(builder, requiredIndexEntries);
            stream.flush();
            recordPosition(rowIndexPosition);
        }

        @Override
        void recordPosition(PositionRecorder recorder) throws IOException {
            super.recordPosition(recorder);
            stream.getPosition(recorder);
        }
    }

    private static class StringTreeWriter extends TreeWriter {
        private final PositionedOutputStream stringOutput;
        private final RunLengthIntegerWriter lengthOutput;
        private final PositionedOutputStream inDictionaryStream;
        private final BitFieldWriter inDictionary;
        private StringDictionaryEncoder dictionary;
        private DynamicIntArray rows;
        private final RunLengthIntegerWriter directLengthOutput;
        private final RunLengthIntegerWriter strideDictionaryLengthOutput;
        private final List<OrcProto.RowIndexEntry> savedRowIndex = new ArrayList<OrcProto.RowIndexEntry>();
        private final boolean buildIndex;
        private final List<Long> rowIndexValueCount = new ArrayList<Long>();
        private final StreamFactory writer;
        // If the number of keys in a dictionary is greater than this fraction of the total number of
        // non-null rows, turn off dictionary encoding
        private final float dictionaryKeySizeThreshold;
        // If the number of keys in a dictionary is greater than this fraction of the total number of
        // non-null rows, don't use the estimated entropy heuristic to turn off dictionary encoding
        private final float entropyKeySizeThreshold;
        private final int entropyMinSamples;
        private final float entropyDictSampleFraction;
        private final int entropyThreshold;

        private boolean useDictionaryEncoding = true;
        private final boolean useStrideDictionaries;
        private final boolean sortKeys;

        private final Text[] buffer;
        private int bufferIndex = 0;
        private long bufferedBytes = 0;
        private final int recomputeStripeEncodingInterval;
        private PositionedOutputStream rowOutput;
        private final PositionedOutputStream strideDictionaryOutput;
        private boolean abandonDictionaries = false;
        private int dictionarySize;

        StringTreeWriter(int columnId, ObjectInspector inspector, StreamFactory writerFactory, boolean nullable,
                Configuration conf, boolean useVInts, boolean lowMemoryMode, MemoryEstimate memoryEstimate)
                throws IOException {
            super(columnId, inspector, writerFactory, nullable, conf, useVInts, memoryEstimate);
            writer = writerFactory;
            sortKeys = OrcConf.getBoolVar(conf, OrcConf.ConfVars.HIVE_ORC_DICTIONARY_SORT_KEYS);
            useStrideDictionaries = OrcConf.getBoolVar(conf, OrcConf.ConfVars.HIVE_ORC_BUILD_STRIDE_DICTIONARY);
            recomputeStripeEncodingInterval = OrcConf.getIntVar(conf,
                    OrcConf.ConfVars.HIVE_ORC_DICTIONARY_ENCODING_INTERVAL);

            if (!lowMemoryMode) {
                dictionary = new StringDictionaryEncoder(sortKeys, useStrideDictionaries, memoryEstimate);
                rows = new DynamicIntArray(memoryEstimate);
            } else {
                abandonDictionaries = true;
                rowOutput = writer.createStream(id, OrcProto.Stream.Kind.DATA);
                useDictionaryEncoding = false;
            }
            stringOutput = writer.createStream(id, OrcProto.Stream.Kind.DICTIONARY_DATA);
            lengthOutput = new RunLengthIntegerWriter(writer.createStream(id, OrcProto.Stream.Kind.LENGTH), false,
                    INT_BYTE_SIZE, useVInts);
            inDictionaryStream = writer.createStream(id, OrcProto.Stream.Kind.IN_DICTIONARY);
            inDictionary = new BitFieldWriter(inDictionaryStream, 1);
            strideDictionaryLengthOutput = new RunLengthIntegerWriter(
                    writer.createStream(id, OrcProto.Stream.Kind.STRIDE_DICTIONARY_LENGTH), false, INT_BYTE_SIZE,
                    useVInts);
            strideDictionaryOutput = writer.createStream(id, OrcProto.Stream.Kind.STRIDE_DICTIONARY);
            directLengthOutput = new RunLengthIntegerWriter(writer.createStream(id, OrcProto.Stream.Kind.LENGTH),
                    false, INT_BYTE_SIZE, useVInts);
            dictionaryKeySizeThreshold = OrcConf.getFloatVar(conf,
                    OrcConf.ConfVars.HIVE_ORC_DICTIONARY_STRING_KEY_SIZE_THRESHOLD);
            entropyKeySizeThreshold = OrcConf.getFloatVar(conf,
                    OrcConf.ConfVars.HIVE_ORC_ENTROPY_KEY_STRING_SIZE_THRESHOLD);
            entropyMinSamples = OrcConf.getIntVar(conf, OrcConf.ConfVars.HIVE_ORC_ENTROPY_STRING_MIN_SAMPLES);
            entropyDictSampleFraction = OrcConf.getFloatVar(conf,
                    OrcConf.ConfVars.HIVE_ORC_ENTROPY_STRING_DICT_SAMPLE_FRACTION);
            entropyThreshold = OrcConf.getIntVar(conf, OrcConf.ConfVars.HIVE_ORC_ENTROPY_STRING_THRESHOLD);

            int bufferLength = OrcConf.getIntVar(conf, OrcConf.ConfVars.HIVE_ORC_ROW_BUFFER_SIZE);
            buffer = new Text[bufferLength];

            recordPosition(rowIndexPosition);
            rowIndexValueCount.add(0L);
            buildIndex = writer.buildIndex();
            if (buildIndex && lowMemoryMode) {
                rowOutput.getPosition(rowIndexPosition);
                directLengthOutput.getPosition(rowIndexPosition);
            }
        }

        boolean determineEncodingStripe() {
            return (getNumStripes() % recomputeStripeEncodingInterval) == 0 && !abandonDictionaries;
        }

        @Override
        void write(Object obj) throws IOException {
            if (obj != null) {
                Text val = ((StringObjectInspector) inspector).getPrimitiveWritableObject(obj);
                buffer[bufferIndex++] = new Text(val);
                setRawDataSize(val.getLength());
                // Increment the memory estimate by the buffered bytes
                memoryEstimate.incrementTotalMemory(val.getLength());
                bufferedBytes += val.getLength();
            } else {
                buffer[bufferIndex++] = null;
                setRawDataSize(RawDatasizeConst.NULL_SIZE);
            }
            if (bufferIndex == buffer.length) {
                flush();
            }
        }

        @Override
        void flush() throws IOException {
            for (int i = 0; i < bufferIndex; i++) {
                Text val = buffer[i];
                // Make sure we don't end up storing it twice
                buffer[i] = null;
                if (val != null) {
                    indexStatistics.updateString(val.toString());
                    if (useCarriedOverDirectEncoding()) {
                        rowOutput.write(val.getBytes());
                        directLengthOutput.write(val.getLength());
                    } else {
                        rows.add(dictionary.add(val, savedRowIndex.size()));
                    }
                }
                super.flushRow(val);
            }
            bufferIndex = 0;
            memoryEstimate.decrementTotalMemory(bufferedBytes);
            bufferedBytes = 0;
        }

        private boolean isEntropyThresholdExceeded(Set<Character> chars, Text text, int index) {
            dictionary.getText(text, index);
            for (char character : text.toString().toCharArray()) {
                chars.add(character);
            }
            return chars.size() > entropyThreshold;
        }

        private int[] getSampleIndecesForEntropy() {
            int numSamples = Math.max(entropyMinSamples, (int) (entropyDictSampleFraction * dictionary.size()));
            int[] indeces = new int[dictionary.size()];
            int[] samples = new int[numSamples];
            Random rand = new Random();

            // The goal of this loop is to select numSamples number of distinct indeces of
            // dictionary
            //
            // The loop works as follows, start with an array of zeros
            // On each iteration pick a random number in the range of 0 to size of the dictionary
            // minus one minus the number of previous iterations, thus the effective size of the
            // array is decreased by one with each iteration (not actually but logically)
            // Look at the value of the array at that random index, if it is 0, the sample index is
            // the random index because we've never looked at this position before, if it's
            // nonzero, that value is the sample index
            // Then take the value at the end of the logical size of the array (size of the
            // dictionary minus one minus the number of iterations) if it's 0 put that index into
            // the array at the random index, otherwise put the nonzero value
            // Thus, by reducing the logical size of the array and moving the value at the end of
            // the array we are removing indexes we have previously visited and making sure we do
            // not lose any indexes
            for (int i = 0; i < numSamples; i++) {
                int index = rand.nextInt(dictionary.size() - i);
                if (indeces[index] == 0) {
                    samples[i] = index;
                } else {
                    samples[i] = indeces[index];
                }
                indeces[index] = indeces[dictionary.size() - i - 1] == 0 ? dictionary.size() - i - 1
                        : indeces[dictionary.size() - i - 1];
            }

            return samples;
        }

        private boolean useDictionaryEncodingEntropyHeuristic() {
            Set<Character> chars = new HashSet<Character>();
            Text text = new Text();
            if (dictionary.size() > entropyMinSamples) {

                int[] samples = getSampleIndecesForEntropy();

                for (int sampleIndex : samples) {
                    if (isEntropyThresholdExceeded(chars, text, sampleIndex)) {
                        return true;
                    }
                }
            } else {
                for (int i = 0; i < dictionary.size(); i++) {
                    if (isEntropyThresholdExceeded(chars, text, i)) {
                        return true;
                    }
                }
            }

            return false;
        }

        /**
         * Returns true iff the encoding is not being determined using this stripe, and
         * the previously determined encoding was direct.
         */
        private boolean useCarriedOverDirectEncoding() {
            return !determineEncodingStripe() && !useDictionaryEncoding;
        }

        @Override
        void writeStripe(OrcProto.StripeFooter.Builder builder, int requiredIndexEntries) throws IOException {

            if (determineEncodingStripe()) {

                // Set the flag indicating whether or not to use dictionary encoding based on whether
                // or not the fraction of distinct keys over number of non-null rows is less than the
                // configured threshold, and whether or not the number of distinct characters in a sample
                // of entries in the dictionary (the estimated entropy) exceeds the configured threshold
                if (rows.size() > 0) {
                    useDictionaryEncoding = true;
                    // The fraction of non-null values in this column that are repeats of values in the
                    // dictionary
                    float repeatedValuesFraction = (float) (rows.size() - dictionary.size()) / (float) rows.size();

                    // If the number of repeated values is small enough, consider using the entropy heuristic
                    // If the number of repeated values is high, even in the presence of low entropy,
                    // dictionary encoding can provide benefits beyond just zlib
                    if (repeatedValuesFraction <= entropyKeySizeThreshold) {
                        useDictionaryEncoding = useDictionaryEncodingEntropyHeuristic();
                    }

                    // dictionaryKeySizeThreshold is the fraction of keys that are distinct beyond
                    // which dictionary encoding is turned off
                    // so 1 - dictionaryKeySizeThreshold is the number of repeated values below which
                    // dictionary encoding should be turned off
                    useDictionaryEncoding = useDictionaryEncoding
                            && (repeatedValuesFraction > 1.0 - dictionaryKeySizeThreshold);
                }
            }

            if (useDictionaryEncoding) {
                rowOutput = new RunLengthIntegerWriter(writer.createStream(id, OrcProto.Stream.Kind.DATA), false,
                        INT_BYTE_SIZE, useVInts);
            } else if (determineEncodingStripe()) {
                rowOutput = writer.createStream(id, OrcProto.Stream.Kind.DATA);
            }

            final int[] dumpOrder;
            final int[] counts;
            final int[] strideDictionarySizes;
            final boolean[] strideDictionaryIndexPopulated;

            if (useDictionaryEncoding) {
                dumpOrder = new int[dictionary.size()];
                counts = new int[dictionary.size()];
                strideDictionarySizes = new int[savedRowIndex.size()];
                strideDictionaryIndexPopulated = new boolean[savedRowIndex.size()];
                OrcProto.RowIndexEntry.Builder base = savedRowIndex.get(0).toBuilder();
                PositionRecorder recorder = new RowIndexPositionRecorder(base);
                strideDictionaryOutput.getPosition(recorder);
                strideDictionaryLengthOutput.getPosition(recorder);
                savedRowIndex.set(0, base.build());
                strideDictionaryIndexPopulated[0] = true;

                // Traverse the red-black tree writing out the bytes and lengths; and
                // creating the map from the original order to the final sorted order.
                dictionary.visit(new StringDictionaryEncoder.Visitor<Text>() {
                    private int currentId = 0;
                    private int directId = 0;
                    private int previousIndex = 0;

                    @Override
                    public void visit(StringDictionaryEncoder.VisitorContext<Text> context) throws IOException {
                        counts[context.getOriginalPosition()] = context.getCount();
                        if (!useStrideDictionaries || context.getCount() > 1) {
                            dictionarySize++;
                            context.writeBytes(stringOutput);
                            lengthOutput.write(context.getLength());
                            dumpOrder[context.getOriginalPosition()] = currentId++;
                        } else {
                            int nextIndex = context.getIndexStride();
                            if (nextIndex != previousIndex) {
                                for (int i = previousIndex; i < nextIndex; i++) {
                                    OrcProto.RowIndexEntry.Builder base = savedRowIndex.get(i + 1).toBuilder();
                                    PositionRecorder recorder = new RowIndexPositionRecorder(base);
                                    strideDictionaryOutput.getPosition(recorder);
                                    strideDictionaryLengthOutput.getPosition(recorder);
                                    savedRowIndex.set(i + 1, base.build());
                                    strideDictionaryIndexPopulated[i + 1] = true;
                                }

                                previousIndex = context.getIndexStride();
                                directId = 0;
                            }
                            context.writeBytes(strideDictionaryOutput);
                            strideDictionarySizes[previousIndex]++;
                            strideDictionaryLengthOutput.write(context.getLength());
                            dumpOrder[context.getOriginalPosition()] = directId++;
                        }
                    }
                });
            } else {
                dumpOrder = null;
                counts = null;
                strideDictionarySizes = null;
                strideDictionaryIndexPopulated = null;
            }

            if (!useCarriedOverDirectEncoding()) {
                writeData(useDictionaryEncoding, dumpOrder, counts, strideDictionarySizes,
                        strideDictionaryIndexPopulated);
            }
            // we need to build the rowindex before calling super, since it
            // writes it out.
            super.writeStripe(builder, requiredIndexEntries);
            rowOutput.flush();
            if (useDictionaryEncoding) {
                stringOutput.unsuppress();
                lengthOutput.unsuppress();
                inDictionary.flush();
                strideDictionaryOutput.flush();
                strideDictionaryLengthOutput.flush();
                if (dictionarySize == dictionary.size()) {
                    inDictionaryStream.suppress();
                    strideDictionaryOutput.suppress();
                    strideDictionaryLengthOutput.suppress();
                } else {
                    inDictionaryStream.unsuppress();
                    strideDictionaryOutput.unsuppress();
                    strideDictionaryLengthOutput.unsuppress();
                }
                directLengthOutput.suppress();
                stringOutput.flush();
                lengthOutput.flush();
            } else {
                directLengthOutput.unsuppress();
                stringOutput.suppress();
                lengthOutput.suppress();
                inDictionaryStream.suppress();
                strideDictionaryOutput.suppress();
                strideDictionaryLengthOutput.suppress();
                directLengthOutput.flush();
            }

            // reset all of the fields to be ready for the next stripe.
            savedRowIndex.clear();
            rowIndexValueCount.clear();
            recordPosition(rowIndexPosition);
            dictionarySize = 0;
            if (useCarriedOverDirectEncoding()) {
                rowOutput = writer.createStream(id, OrcProto.Stream.Kind.DATA);
                rowOutput.getPosition(rowIndexPosition);
                directLengthOutput.getPosition(rowIndexPosition);
                if (dictionary != null) {
                    dictionary.cleanup();
                    dictionary = null;
                    rows.cleanup();
                    rows = null;
                }
            } else {
                if (dictionary == null) {
                    dictionary = new StringDictionaryEncoder(sortKeys, useStrideDictionaries, memoryEstimate);
                } else {
                    dictionary.clear();
                }
                if (rows == null) {
                    rows = new DynamicIntArray(memoryEstimate);
                } else {
                    rows.clear();
                }
            }
            rowIndexValueCount.add(0L);

        }

        private void convertDictionaryToDirect() throws IOException {
            // If this is still in the first index stride, savedRowIndex is empty, so we need to
            // explicitly record the positions for the first index stride
            writeData(false, null, null, null, null);
        }

        private void writeData(boolean useDictionaryEncoding, int[] dumpOrder, int[] counts,
                int[] strideDictionarySizes, boolean[] strideDictionaryIndexPopulated) throws IOException {
            int rowIndexEntry = 0;
            OrcProto.RowIndex.Builder rowIndex = getRowIndex();
            int length = rows.size();
            Text text = new Text();
            for (int i = 0; i <= length; ++i) {
                // now that we are writing out the row values, we can finalize the
                // row index
                if (buildIndex) {
                    // If we are not using dictionary encoding and savedRowIndex is not empty, this means
                    // the conversion from dictionary encoding to direct encoding, so allow rowIndexEntry to
                    // exceed the number of saved row indeces, to create the index entry for the current
                    //stride
                    while (rowIndexEntry < rowIndexValueCount.size() && i == rowIndexValueCount.get(rowIndexEntry)
                            && (rowIndexEntry < savedRowIndex.size()
                                    || (!useDictionaryEncoding && rowIndexEntry == savedRowIndex.size()))) {
                        OrcProto.RowIndexEntry.Builder base = null;
                        RowIndexPositionRecorder recorder;
                        if (rowIndexEntry < savedRowIndex.size()) {
                            base = savedRowIndex.get(rowIndexEntry).toBuilder();
                            recorder = new RowIndexPositionRecorder(base);
                        } else {
                            recorder = rowIndexPosition;
                        }

                        if (useStrideDictionaries && useDictionaryEncoding
                                && !strideDictionaryIndexPopulated[rowIndexEntry]
                                && dictionary.size() != dictionarySize) {
                            strideDictionaryOutput.getPosition(recorder);
                            strideDictionaryLengthOutput.getPosition(recorder);
                        }
                        if (useStrideDictionaries && strideDictionarySizes != null
                                && dictionary.size() != dictionarySize) {
                            base.addPositions(strideDictionarySizes[rowIndexEntry]);
                        }
                        recordOutputPosition(rowOutput, recorder);
                        if (rowIndexEntry < savedRowIndex.size()) {
                            // If we are constructing an index entry from a saved row index, add it
                            rowIndex.addEntry(base.build());
                        }
                        rowIndexEntry++;
                    }
                }
                if (i != length) {
                    if (useDictionaryEncoding) {
                        rowOutput.write(dumpOrder[rows.get(i)]);
                        if (!useStrideDictionaries || counts[rows.get(i)] > 1) {
                            inDictionary.write(1);
                        } else {
                            inDictionary.write(0);
                        }
                    } else {
                        dictionary.getText(text, rows.get(i));
                        rowOutput.write(text.getBytes(), 0, text.getLength());
                        directLengthOutput.write(text.getLength());
                    }
                }
            }
        }

        // Calls getPosition on the row output stream if dictionary encoding is used, and the direct
        // output stream if direct encoding is used
        private void recordOutputPosition(PositionedOutputStream rowOutput, RowIndexPositionRecorder recorder)
                throws IOException {
            rowOutput.getPosition(recorder);
            if (!useDictionaryEncoding) {
                directLengthOutput.getPosition(recorder);
            } else if (dictionary.size() != dictionarySize) {
                inDictionary.getPosition(recorder);
            }
        }

        @Override
        OrcProto.ColumnEncoding getEncoding() {
            // Returns the encoding used for the last call to writeStripe
            if (useDictionaryEncoding) {
                return OrcProto.ColumnEncoding.newBuilder().setKind(OrcProto.ColumnEncoding.Kind.DICTIONARY)
                        .setDictionarySize(dictionarySize).build();
            } else {
                return OrcProto.ColumnEncoding.newBuilder().setKind(OrcProto.ColumnEncoding.Kind.DIRECT).build();
            }
        }

        /**
         * This method doesn't call the super method, because unlike most of the
         * other TreeWriters, this one can't record the position in the streams
         * until the stripe is being flushed. Therefore it saves all of the entries
         * and augments them with the final information as the stripe is written.
         * @throws IOException
         */
        @Override
        void createRowIndexEntry() throws IOException {
            getFileStatistics().merge(indexStatistics);
            OrcProto.RowIndexEntry.Builder rowIndexEntry = getRowIndexEntry();
            rowIndexEntry.setStatistics(indexStatistics.serialize());
            indexStatistics.reset();
            if (useCarriedOverDirectEncoding()) {
                getRowIndex().addEntry(rowIndexEntry);
            } else {
                savedRowIndex.add(rowIndexEntry.build());
            }
            rowIndexEntry.clear();
            recordPosition(rowIndexPosition);
            if (useCarriedOverDirectEncoding()) {
                rowOutput.getPosition(rowIndexPosition);
                directLengthOutput.getPosition(rowIndexPosition);
            } else {
                rowIndexValueCount.add(Long.valueOf(rows.size()));
            }
        }

        @Override
        public void abandonDictionaries() throws IOException {
            boolean useCarriedOverDirectEncoding = useCarriedOverDirectEncoding();
            abandonDictionaries = true;
            if (!useCarriedOverDirectEncoding) {
                rowOutput = writer.createStream(id, OrcProto.Stream.Kind.DATA);
                useDictionaryEncoding = false;
                convertDictionaryToDirect();
                if (rows.size() == 0) {
                    rowOutput.getPosition(rowIndexPosition);
                    directLengthOutput.getPosition(rowIndexPosition);
                }
            }
            if (dictionary != null) {
                dictionary.cleanup();
                dictionary = null;
                rows.cleanup();
                rows = null;
            }
            savedRowIndex.clear();
        }
    }

    private static class BinaryTreeWriter extends TreeWriter {
        private final PositionedOutputStream stream;
        private final RunLengthIntegerWriter length;

        BinaryTreeWriter(int columnId, ObjectInspector inspector, StreamFactory writer, boolean nullable,
                Configuration conf, boolean useVInts, boolean lowMemoryMode, MemoryEstimate memoryEstimate)
                throws IOException {
            super(columnId, inspector, writer, nullable, conf, useVInts, memoryEstimate);
            this.stream = writer.createStream(id, OrcProto.Stream.Kind.DATA);
            this.length = new RunLengthIntegerWriter(writer.createStream(id, OrcProto.Stream.Kind.LENGTH), false,
                    INT_BYTE_SIZE, useVInts);
            recordPosition(rowIndexPosition);
        }

        @Override
        void write(Object obj) throws IOException {
            long rawDataSize = 0;
            if (obj != null) {
                BytesWritable val = ((BinaryObjectInspector) inspector).getPrimitiveWritableObject(obj);
                stream.write(val.getBytes(), 0, val.getLength());
                length.write(val.getLength());

                // Raw data size is the length of the BytesWritable, i.e. the number of bytes
                rawDataSize = val.getLength();
            }
            super.write(obj, rawDataSize);
        }

        @Override
        void writeStripe(OrcProto.StripeFooter.Builder builder, int requiredIndexEntries) throws IOException {
            super.writeStripe(builder, requiredIndexEntries);
            stream.flush();
            length.flush();
            recordPosition(rowIndexPosition);
        }

        @Override
        void recordPosition(PositionRecorder recorder) throws IOException {
            super.recordPosition(recorder);
            stream.getPosition(recorder);
            length.getPosition(recorder);
        }
    }

    public static final int MILLIS_PER_SECOND = 1000;
    public static final long BASE_TIMESTAMP = Timestamp.valueOf("2015-01-01 00:00:00").getTime()
            / MILLIS_PER_SECOND;

    private static class TimestampTreeWriter extends TreeWriter {
        private final RunLengthIntegerWriter seconds;
        private final RunLengthIntegerWriter nanos;

        TimestampTreeWriter(int columnId, ObjectInspector inspector, StreamFactory writer, boolean nullable,
                Configuration conf, boolean useVInts, boolean lowMemoryMode, MemoryEstimate memoryEstimate)
                throws IOException {
            super(columnId, inspector, writer, nullable, conf, useVInts, memoryEstimate);
            this.seconds = new RunLengthIntegerWriter(writer.createStream(id, OrcProto.Stream.Kind.DATA), true,
                    LONG_BYTE_SIZE, useVInts);
            this.nanos = new RunLengthIntegerWriter(writer.createStream(id, OrcProto.Stream.Kind.NANO_DATA), false,
                    LONG_BYTE_SIZE, useVInts);
            recordPosition(rowIndexPosition);
        }

        @Override
        void write(Object obj) throws IOException {
            // Raw data size is:
            //   the number of bytes needed to store the milliseconds since the epoch
            //   (8 since it's a long)
            //   +
            //   the number of bytes needed to store the nanos field (4 since it's an int)
            super.write(obj, RawDatasizeConst.TIMESTAMP_SIZE);
            if (obj != null) {
                Timestamp val = ((TimestampObjectInspector) inspector).getPrimitiveJavaObject(obj);
                seconds.write((val.getTime() / MILLIS_PER_SECOND) - BASE_TIMESTAMP);
                nanos.write(formatNanos(val.getNanos()));
            }
        }

        @Override
        void writeStripe(OrcProto.StripeFooter.Builder builder, int requiredIndexEntries) throws IOException {
            super.writeStripe(builder, requiredIndexEntries);
            seconds.flush();
            nanos.flush();
            recordPosition(rowIndexPosition);
        }

        private static long formatNanos(int nanos) {
            if (nanos == 0) {
                return 0;
            } else if (nanos % 100 != 0) {
                return ((long) nanos) << 3;
            } else {
                nanos /= 100;
                int trailingZeros = 1;
                while (nanos % 10 == 0 && trailingZeros < 7) {
                    nanos /= 10;
                    trailingZeros += 1;
                }
                return ((long) nanos) << 3 | trailingZeros;
            }
        }

        @Override
        void recordPosition(PositionRecorder recorder) throws IOException {
            super.recordPosition(recorder);
            seconds.getPosition(recorder);
            nanos.getPosition(recorder);
        }
    }

    private static class StructTreeWriter extends TreeWriter {
        private final List<? extends StructField> fields;

        StructTreeWriter(int columnId, ObjectInspector inspector, StreamFactory writer, boolean nullable,
                Configuration conf, boolean useVInts, boolean lowMemoryMode, MemoryEstimate memoryEstimate)
                throws IOException {
            super(columnId, inspector, writer, nullable, conf, useVInts, memoryEstimate);
            StructObjectInspector structObjectInspector = (StructObjectInspector) inspector;
            fields = structObjectInspector.getAllStructFieldRefs();
            childrenWriters = new TreeWriter[fields.size()];
            for (int i = 0; i < childrenWriters.length; ++i) {
                childrenWriters[i] = createTreeWriter(fields.get(i).getFieldObjectInspector(), writer, true, conf,
                        useVInts, lowMemoryMode, memoryEstimate);
            }
            recordPosition(rowIndexPosition);
        }

        @Override
        void write(Object obj) throws IOException {
            long rawDataSize = 0;
            if (obj != null) {
                StructObjectInspector insp = (StructObjectInspector) inspector;
                List<Object> fieldDataList = insp.getStructFieldsDataAsList(obj);

                for (int i = 0; i < fields.size(); ++i) {
                    TreeWriter writer = childrenWriters[i];
                    writer.write(fieldDataList.get(i));
                    rawDataSize += writer.getRowRawDataSize();
                }
            }
            super.write(obj, rawDataSize);
        }

        @Override
        void writeStripe(OrcProto.StripeFooter.Builder builder, int requiredIndexEntries) throws IOException {
            super.writeStripe(builder, requiredIndexEntries);
            for (TreeWriter child : childrenWriters) {
                child.writeStripe(builder, requiredIndexEntries);
            }
            recordPosition(rowIndexPosition);
        }
    }

    private static class ListTreeWriter extends TreeWriter {
        private final RunLengthIntegerWriter lengths;

        ListTreeWriter(int columnId, ObjectInspector inspector, StreamFactory writer, boolean nullable,
                Configuration conf, boolean useVInts, boolean lowMemoryMode, MemoryEstimate memoryEstimate)
                throws IOException {
            super(columnId, inspector, writer, nullable, conf, useVInts, memoryEstimate);
            ListObjectInspector listObjectInspector = (ListObjectInspector) inspector;
            childrenWriters = new TreeWriter[1];
            childrenWriters[0] = createTreeWriter(listObjectInspector.getListElementObjectInspector(), writer, true,
                    conf, useVInts, lowMemoryMode, memoryEstimate);
            lengths = new RunLengthIntegerWriter(writer.createStream(columnId, OrcProto.Stream.Kind.LENGTH), false,
                    INT_BYTE_SIZE, useVInts);
            recordPosition(rowIndexPosition);
        }

        @Override
        void write(Object obj) throws IOException {
            long rawDataSize = 0;
            if (obj != null) {
                ListObjectInspector insp = (ListObjectInspector) inspector;
                int len = insp.getListLength(obj);
                lengths.write(len);
                for (int i = 0; i < len; ++i) {
                    childrenWriters[0].write(insp.getListElement(obj, i));
                    rawDataSize += childrenWriters[0].getRowRawDataSize();
                }
            }
            super.write(obj, rawDataSize);
        }

        @Override
        void writeStripe(OrcProto.StripeFooter.Builder builder, int requiredIndexEntries) throws IOException {
            super.writeStripe(builder, requiredIndexEntries);
            lengths.flush();
            for (TreeWriter child : childrenWriters) {
                child.writeStripe(builder, requiredIndexEntries);
            }
            recordPosition(rowIndexPosition);
        }

        @Override
        void recordPosition(PositionRecorder recorder) throws IOException {
            super.recordPosition(recorder);
            lengths.getPosition(recorder);
        }
    }

    private static class MapTreeWriter extends TreeWriter {
        private final RunLengthIntegerWriter lengths;

        MapTreeWriter(int columnId, ObjectInspector inspector, StreamFactory writer, boolean nullable,
                Configuration conf, boolean useVInts, boolean lowMemoryMode, MemoryEstimate memoryEstimate)
                throws IOException {
            super(columnId, inspector, writer, nullable, conf, useVInts, memoryEstimate);
            MapObjectInspector insp = (MapObjectInspector) inspector;
            childrenWriters = new TreeWriter[2];
            childrenWriters[0] = createTreeWriter(insp.getMapKeyObjectInspector(), writer, true, conf, useVInts,
                    lowMemoryMode, memoryEstimate);
            childrenWriters[1] = createTreeWriter(insp.getMapValueObjectInspector(), writer, true, conf, useVInts,
                    lowMemoryMode, memoryEstimate);
            lengths = new RunLengthIntegerWriter(writer.createStream(columnId, OrcProto.Stream.Kind.LENGTH), false,
                    INT_BYTE_SIZE, useVInts);
            recordPosition(rowIndexPosition);
        }

        @Override
        void write(Object obj) throws IOException {
            long rawDataSize = 0;
            if (obj != null) {
                MapObjectInspector insp = (MapObjectInspector) inspector;
                // this sucks, but it will have to do until we can get a better
                // accessor in the MapObjectInspector.
                Map<?, ?> valueMap = insp.getMap(obj);
                // Don't use getMapSize(), it's inconsistent for some object inspectors
                int len = valueMap.size();
                lengths.write(len);
                for (Map.Entry<?, ?> entry : valueMap.entrySet()) {
                    childrenWriters[0].write(entry.getKey());
                    childrenWriters[1].write(entry.getValue());
                    rawDataSize += childrenWriters[0].getRowRawDataSize();
                    rawDataSize += childrenWriters[1].getRowRawDataSize();
                }
            }
            super.write(obj, rawDataSize);
        }

        @Override
        void writeStripe(OrcProto.StripeFooter.Builder builder, int requiredIndexEntries) throws IOException {
            super.writeStripe(builder, requiredIndexEntries);
            lengths.flush();
            for (TreeWriter child : childrenWriters) {
                child.writeStripe(builder, requiredIndexEntries);
            }
            recordPosition(rowIndexPosition);
        }

        @Override
        void recordPosition(PositionRecorder recorder) throws IOException {
            super.recordPosition(recorder);
            lengths.getPosition(recorder);
        }
    }

    private static class UnionTreeWriter extends TreeWriter {
        private final RunLengthByteWriter tags;

        UnionTreeWriter(int columnId, ObjectInspector inspector, StreamFactory writer, boolean nullable,
                Configuration conf, boolean useVInts, boolean lowMemoryMode, MemoryEstimate memoryEstimate)
                throws IOException {
            super(columnId, inspector, writer, nullable, conf, useVInts, memoryEstimate);
            UnionObjectInspector insp = (UnionObjectInspector) inspector;
            List<ObjectInspector> choices = insp.getObjectInspectors();
            childrenWriters = new TreeWriter[choices.size()];
            for (int i = 0; i < childrenWriters.length; ++i) {
                childrenWriters[i] = createTreeWriter(choices.get(i), writer, true, conf, useVInts, lowMemoryMode,
                        memoryEstimate);
            }
            tags = new RunLengthByteWriter(writer.createStream(columnId, OrcProto.Stream.Kind.DATA));
            recordPosition(rowIndexPosition);
        }

        @Override
        void write(Object obj) throws IOException {
            long rawDataSize = 0;
            if (obj != null) {
                UnionObjectInspector insp = (UnionObjectInspector) inspector;
                byte tag = insp.getTag(obj);
                tags.write(tag);
                childrenWriters[tag].write(insp.getField(obj));
                // raw data size is size of tag (1) + size of value
                rawDataSize = childrenWriters[tag].getRowRawDataSize() + RawDatasizeConst.UNION_TAG_SIZE;
            }
            super.write(obj, rawDataSize);
        }

        @Override
        void writeStripe(OrcProto.StripeFooter.Builder builder, int requiredIndexEntries) throws IOException {
            super.writeStripe(builder, requiredIndexEntries);
            tags.flush();
            for (TreeWriter child : childrenWriters) {
                child.writeStripe(builder, requiredIndexEntries);
            }
            recordPosition(rowIndexPosition);
        }

        @Override
        void recordPosition(PositionRecorder recorder) throws IOException {
            super.recordPosition(recorder);
            tags.getPosition(recorder);
        }
    }

    private static TreeWriter createTreeWriter(ObjectInspector inspector, StreamFactory streamFactory,
            boolean nullable, Configuration conf, boolean useVInts, boolean lowMemoryMode,
            MemoryEstimate memoryEstimate) throws IOException {
        switch (inspector.getCategory()) {
        case PRIMITIVE:
            switch (((PrimitiveObjectInspector) inspector).getPrimitiveCategory()) {
            case BOOLEAN:
                return new BooleanTreeWriter(streamFactory.getNextColumnId(), inspector, streamFactory, nullable,
                        conf, useVInts, lowMemoryMode, memoryEstimate);
            case BYTE:
                return new ByteTreeWriter(streamFactory.getNextColumnId(), inspector, streamFactory, nullable, conf,
                        useVInts, lowMemoryMode, memoryEstimate);
            case SHORT:
                return new IntegerTreeWriter(streamFactory.getNextColumnId(), inspector, streamFactory, nullable,
                        conf, useVInts, SHORT_BYTE_SIZE, lowMemoryMode, memoryEstimate);
            case INT:
                return new IntegerTreeWriter(streamFactory.getNextColumnId(), inspector, streamFactory, nullable,
                        conf, useVInts, INT_BYTE_SIZE, lowMemoryMode, memoryEstimate);
            case LONG:
                return new IntegerTreeWriter(streamFactory.getNextColumnId(), inspector, streamFactory, nullable,
                        conf, useVInts, LONG_BYTE_SIZE, lowMemoryMode, memoryEstimate);
            case FLOAT:
                return new FloatTreeWriter(streamFactory.getNextColumnId(), inspector, streamFactory, nullable,
                        conf, useVInts, lowMemoryMode, memoryEstimate);
            case DOUBLE:
                return new DoubleTreeWriter(streamFactory.getNextColumnId(), inspector, streamFactory, nullable,
                        conf, useVInts, lowMemoryMode, memoryEstimate);
            case STRING:
                return new StringTreeWriter(streamFactory.getNextColumnId(), inspector, streamFactory, nullable,
                        conf, useVInts, lowMemoryMode, memoryEstimate);
            case BINARY:
                return new BinaryTreeWriter(streamFactory.getNextColumnId(), inspector, streamFactory, nullable,
                        conf, useVInts, lowMemoryMode, memoryEstimate);
            case TIMESTAMP:
                return new TimestampTreeWriter(streamFactory.getNextColumnId(), inspector, streamFactory, nullable,
                        conf, useVInts, lowMemoryMode, memoryEstimate);
            default:
                throw new IllegalArgumentException(
                        "Bad primitive category " + ((PrimitiveObjectInspector) inspector).getPrimitiveCategory());
            }
        case STRUCT:
            return new StructTreeWriter(streamFactory.getNextColumnId(), inspector, streamFactory, nullable, conf,
                    useVInts, lowMemoryMode, memoryEstimate);
        case MAP:
            return new MapTreeWriter(streamFactory.getNextColumnId(), inspector, streamFactory, nullable, conf,
                    useVInts, lowMemoryMode, memoryEstimate);
        case LIST:
            return new ListTreeWriter(streamFactory.getNextColumnId(), inspector, streamFactory, nullable, conf,
                    useVInts, lowMemoryMode, memoryEstimate);
        case UNION:
            return new UnionTreeWriter(streamFactory.getNextColumnId(), inspector, streamFactory, nullable, conf,
                    useVInts, lowMemoryMode, memoryEstimate);
        default:
            throw new IllegalArgumentException("Bad category: " + inspector.getCategory());
        }
    }

    private static void writeTypes(OrcProto.Footer.Builder builder, TreeWriter treeWriter) {
        OrcProto.Type.Builder type = OrcProto.Type.newBuilder();
        switch (treeWriter.inspector.getCategory()) {
        case PRIMITIVE:
            switch (((PrimitiveObjectInspector) treeWriter.inspector).getPrimitiveCategory()) {
            case BOOLEAN:
                type.setKind(OrcProto.Type.Kind.BOOLEAN);
                break;
            case BYTE:
                type.setKind(OrcProto.Type.Kind.BYTE);
                break;
            case SHORT:
                type.setKind(OrcProto.Type.Kind.SHORT);
                break;
            case INT:
                type.setKind(OrcProto.Type.Kind.INT);
                break;
            case LONG:
                type.setKind(OrcProto.Type.Kind.LONG);
                break;
            case FLOAT:
                type.setKind(OrcProto.Type.Kind.FLOAT);
                break;
            case DOUBLE:
                type.setKind(OrcProto.Type.Kind.DOUBLE);
                break;
            case STRING:
                type.setKind(OrcProto.Type.Kind.STRING);
                break;
            case BINARY:
                type.setKind(OrcProto.Type.Kind.BINARY);
                break;
            case TIMESTAMP:
                type.setKind(OrcProto.Type.Kind.TIMESTAMP);
                break;
            default:
                throw new IllegalArgumentException("Unknown primitive category: "
                        + ((PrimitiveObjectInspector) treeWriter.inspector).getPrimitiveCategory());
            }
            break;
        case LIST:
            type.setKind(OrcProto.Type.Kind.LIST);
            type.addSubtypes(treeWriter.childrenWriters[0].id);
            break;
        case MAP:
            type.setKind(OrcProto.Type.Kind.MAP);
            type.addSubtypes(treeWriter.childrenWriters[0].id);
            type.addSubtypes(treeWriter.childrenWriters[1].id);
            break;
        case STRUCT:
            type.setKind(OrcProto.Type.Kind.STRUCT);
            for (TreeWriter child : treeWriter.childrenWriters) {
                type.addSubtypes(child.id);
            }
            for (StructField field : ((StructTreeWriter) treeWriter).fields) {
                type.addFieldNames(field.getFieldName());
            }
            break;
        case UNION:
            type.setKind(OrcProto.Type.Kind.UNION);
            for (TreeWriter child : treeWriter.childrenWriters) {
                type.addSubtypes(child.id);
            }
            break;
        default:
            throw new IllegalArgumentException("Unknown category: " + treeWriter.inspector.getCategory());
        }
        builder.addTypes(type);
        for (TreeWriter child : treeWriter.childrenWriters) {
            writeTypes(builder, child);
        }
    }

    private void ensureWriter() throws IOException {
        if (rawWriter == null) {
            rawWriter = fs.create(path, false, HDFS_BUFFER_SIZE, fs.getDefaultReplication(),
                    // Use the largest value less than or equal to Integer.MAX_VALUE which is divisible
                    // by dfsBytesPerChecksum as an upper bound (otherwise HDFS will throw an exception)
                    Math.min(stripeSize * 2L, (Integer.MAX_VALUE / dfsBytesPerChecksum) * dfsBytesPerChecksum));
            rawWriter.writeBytes(OrcFile.MAGIC);
            headerLength = rawWriter.getPos();
            writer = new OutStream("metadata", bufferSize, codec, new DirectStream(rawWriter), memoryEstimate);
            protobufWriter = CodedOutputStream.newInstance(writer);
        }
    }

    private void createRowIndexEntry() throws IOException {
        treeWriter.flush();
        treeWriter.createRowIndexEntry();
        rowsInIndex = 0;
    }

    public void addStripe(StripeInformation si, byte[] data) throws IOException {
        ensureWriter();
        OrcProto.StripeInformation dirEntry = OrcProto.StripeInformation.newBuilder().setOffset(rawWriter.getPos())
                .setIndexLength(si.getIndexLength()).setDataLength(si.getDataLength())
                .setNumberOfRows(si.getNumberOfRows()).setFooterLength(si.getFooterLength())
                .setRawDataSize(si.getRawDataSize()).build();
        stripes.add(dirEntry);
        rowCount += si.getNumberOfRows();
        rawWriter.write(data);
    }

    // Before we write out the streams to disk, do some clean up.  For now this consists of two
    // things:
    // 1) Ignore any suppressed streams, this is done by not including them in the return list
    //    and clearing their contents
    // 2) If a stream has dictionary data, put the length stream before the dictionary data in the
    //    DICTIONARY area.  This is because this more closely matches the order in which we actually
    //    read the data.
    private List<Map.Entry<StreamName, BufferedStream>> cleanUpStreams() throws IOException {
        List<Map.Entry<StreamName, BufferedStream>> streamList = new ArrayList<Map.Entry<StreamName, BufferedStream>>(
                streams.size());
        Map<StreamName, Integer> indexMap = new HashMap<StreamName, Integer>(streams.size());

        int increment = 0;

        for (Map.Entry<StreamName, BufferedStream> pair : streams.entrySet()) {
            if (!pair.getValue().isSuppressed()) {
                StreamName name = pair.getKey();
                if (name.getKind() == Kind.LENGTH) {
                    Integer index = indexMap.get(new StreamName(name.getColumn(), Kind.DICTIONARY_DATA));
                    if (index != null) {
                        streamList.add(index + increment, pair);
                        increment++;
                        continue;
                    }
                }

                indexMap.put(name, new Integer(streamList.size()));
                streamList.add(pair);
            } else {
                pair.getValue().clear();
            }
        }

        return streamList;
    }

    protected void flushStripe() throws IOException {
        ensureWriter();

        ReaderWriterProfiler.start(ReaderWriterProfiler.Counter.ENCODING_TIME);
        treeWriter.flush();

        if (buildIndex && rowsInIndex != 0) {
            createRowIndexEntry();
        }
        ReaderWriterProfiler.end(ReaderWriterProfiler.Counter.ENCODING_TIME);
        if (rowsInStripe != 0) {
            int requiredIndexEntries = rowIndexStride == 0 ? 0
                    : (int) ((rowsInStripe + rowIndexStride - 1) / rowIndexStride);
            OrcProto.StripeFooter.Builder builder = OrcProto.StripeFooter.newBuilder();
            long stripeRawDataSize = treeWriter.getStripeRawDataSize();

            ReaderWriterProfiler.start(ReaderWriterProfiler.Counter.SERIALIZATION_TIME);
            treeWriter.writeStripe(builder, requiredIndexEntries);
            ReaderWriterProfiler.end(ReaderWriterProfiler.Counter.SERIALIZATION_TIME);

            long start = rawWriter.getPos();
            long section = start;
            long indexEnd = start;
            List<Map.Entry<StreamName, BufferedStream>> streamList = cleanUpStreams();
            for (Map.Entry<StreamName, BufferedStream> pair : streamList) {
                BufferedStream stream = pair.getValue();
                stream.flush(true);
                stream.spillTo(rawWriter);
                long end = rawWriter.getPos();
                StreamName name = pair.getKey();
                builder.addStreams(OrcProto.Stream.newBuilder().setColumn(name.getColumn()).setKind(name.getKind())
                        .setLength(end - section).setUseVInts(useVInts));
                section = end;
                if (StreamName.Area.INDEX == name.getArea()) {
                    indexEnd = end;
                }
                stream.clear();
            }
            builder.build().writeTo(protobufWriter);
            protobufWriter.flush();
            writer.flush();
            long end = rawWriter.getPos();
            OrcProto.StripeInformation dirEntry = OrcProto.StripeInformation.newBuilder().setOffset(start)
                    .setIndexLength(indexEnd - start).setDataLength(section - indexEnd)
                    .setNumberOfRows(rowsInStripe).setFooterLength(end - section).setRawDataSize(stripeRawDataSize)
                    .build();
            stripes.add(dirEntry);
            rowCount += rowsInStripe;
            rawDataSize += stripeRawDataSize;
            rowsInStripe = 0;
        }
    }

    private OrcProto.CompressionKind writeCompressionKind(CompressionKind kind) {
        switch (kind) {
        case NONE:
            return OrcProto.CompressionKind.NONE;
        case ZLIB:
            return OrcProto.CompressionKind.ZLIB;
        case SNAPPY:
            return OrcProto.CompressionKind.SNAPPY;
        case LZO:
            return OrcProto.CompressionKind.LZO;
        default:
            throw new IllegalArgumentException("Unknown compression " + kind);
        }
    }

    private int writeFileStatistics(OrcProto.Footer.Builder builder, TreeWriter writer,
            ColumnStatisticsImpl[] columnStats, int column) {
        if (columnStats != null) {
            writer.fileStatistics.merge(columnStats[column++]);
        }
        builder.addStatistics(writer.fileStatistics.serialize());
        for (TreeWriter child : writer.getChildrenWriters()) {
            column = writeFileStatistics(builder, child, columnStats, column);
        }
        return column;
    }

    private int writeFooter(long bodyLength, ColumnStatisticsImpl[] columnStats) throws IOException {
        ensureWriter();
        OrcProto.Footer.Builder builder = OrcProto.Footer.newBuilder();
        builder.setContentLength(bodyLength);
        builder.setHeaderLength(headerLength);
        builder.setNumberOfRows(rowCount);
        builder.setRawDataSize(rawDataSize);
        builder.setRowIndexStride(rowIndexStride);
        // serialize the types
        writeTypes(builder, treeWriter);
        // add the stripe information
        for (OrcProto.StripeInformation stripe : stripes) {
            builder.addStripes(stripe);
        }
        // add the column statistics
        writeFileStatistics(builder, treeWriter, columnStats, 0);
        // add all of the user metadata
        for (Map.Entry<String, ByteString> entry : userMetadata.entrySet()) {
            builder.addMetadata(
                    OrcProto.UserMetadataItem.newBuilder().setName(entry.getKey()).setValue(entry.getValue()));
        }
        long startPosn = rawWriter.getPos();
        builder.build().writeTo(protobufWriter);
        protobufWriter.flush();
        writer.flush();
        return (int) (rawWriter.getPos() - startPosn);
    }

    private int writePostScript(int footerLength) throws IOException {
        OrcProto.PostScript.Builder builder = OrcProto.PostScript.newBuilder()
                .setCompression(writeCompressionKind(compress)).setFooterLength(footerLength);
        if (compress != CompressionKind.NONE) {
            builder.setCompressionBlockSize(bufferSize);
        }
        OrcProto.PostScript ps = builder.build();
        // need to write this uncompressed
        long startPosn = rawWriter.getPos();
        ps.writeTo(rawWriter);
        long length = rawWriter.getPos() - startPosn;
        if (length > 255) {
            throw new IllegalArgumentException("PostScript too large at " + length);
        }
        return (int) length;
    }

    @Override
    public synchronized void addUserMetadata(String name, ByteBuffer value) {
        userMetadata.put(name, ByteString.copyFrom(value));
    }

    @Override
    public void addRow(Object row) throws IOException {

        ReaderWriterProfiler.start(ReaderWriterProfiler.Counter.ENCODING_TIME);
        synchronized (this) {
            treeWriter.write(row);
            rowsInStripe += 1;
            if (buildIndex) {
                rowsInIndex += 1;

                if (rowsInIndex >= rowIndexStride) {
                    createRowIndexEntry();
                }
            }
        }
        if (memoryManager.shouldFlush(memoryEstimate, path, stripeSize, maxDictSize)) {
            flushStripe();
        }
        memoryManager.addedRow();
        ReaderWriterProfiler.end(ReaderWriterProfiler.Counter.ENCODING_TIME);
    }

    @Override
    public long getRowRawDataSize() {
        return treeWriter.getRowRawDataSize();
    }

    @Override
    public void close() throws IOException {
        close(null);
    }

    public void close(ColumnStatisticsImpl[] columnStats) throws IOException {
        // remove us from the memory manager so that we don't get any callbacks
        memoryManager.removeWriter(path);
        // actually close the file
        synchronized (this) {
            LOG.info("Flushing the last stripe of file " + path);
            flushStripe();
            LOG.info("Writing out the footer for file " + path);
            int footerLength = writeFooter(rawWriter.getPos(), columnStats);
            LOG.info("Footer length was " + footerLength + ". Writing it at the end of the file " + path);
            rawWriter.writeByte(writePostScript(footerLength));
            LOG.info("Flushing file " + path);
            rawWriter.flush();
            LOG.info("Closing file " + path);
            rawWriter.close();
            LOG.info("Closed file " + path);
        }
    }
}