/**
* Determines which RGs need to be read, after stripes have been determined. SARG is applied, and
* readState is populated for each stripe accordingly.
*/
private boolean determineRgsToRead(
boolean[] globalIncludes, int rowIndexStride, ArrayList<OrcStripeMetadata> metadata)
throws IOException {
SargApplier sargApp = null;
if (sarg != null && rowIndexStride != 0) {
List<OrcProto.Type> types = fileMetadata.getTypes();
String[] colNamesForSarg =
OrcInputFormat.getSargColumnNames(
columnNames, types, globalIncludes, fileMetadata.isOriginalFormat());
sargApp =
new SargApplier(sarg, colNamesForSarg, rowIndexStride, types, globalIncludes.length);
}
boolean hasAnyData = false;
// readState should have been initialized by this time with an empty array.
for (int stripeIxMod = 0; stripeIxMod < readState.length; ++stripeIxMod) {
int stripeIx = stripeIxMod + stripeIxFrom;
StripeInformation stripe = fileMetadata.getStripes().get(stripeIx);
int rgCount = getRgCount(stripe, rowIndexStride);
boolean[] rgsToRead = null;
if (sargApp != null) {
OrcStripeMetadata stripeMetadata = metadata.get(stripeIxMod);
rgsToRead =
sargApp.pickRowGroups(
stripe,
stripeMetadata.getRowIndexes(),
stripeMetadata.getBloomFilterIndexes(),
true);
}
boolean isNone = rgsToRead == SargApplier.READ_NO_RGS,
isAll = rgsToRead == SargApplier.READ_ALL_RGS;
hasAnyData = hasAnyData || !isNone;
if (DebugUtils.isTraceOrcEnabled()) {
if (isNone) {
LlapIoImpl.LOG.info("SARG eliminated all RGs for stripe " + stripeIx);
} else if (!isAll) {
LlapIoImpl.LOG.info(
"SARG picked RGs for stripe " + stripeIx + ": " + DebugUtils.toString(rgsToRead));
} else {
LlapIoImpl.LOG.info("Will read all " + rgCount + " RGs for stripe " + stripeIx);
}
}
assert isAll || isNone || rgsToRead.length == rgCount;
readState[stripeIxMod] = new boolean[columnIds.size()][];
for (int j = 0; j < columnIds.size(); ++j) {
readState[stripeIxMod][j] =
(isAll || isNone) ? rgsToRead : Arrays.copyOf(rgsToRead, rgsToRead.length);
}
adjustRgMetric(rgCount, rgsToRead, isNone, isAll);
}
return hasAnyData;
}
/** Reads the metadata for all stripes in the file. */
private ArrayList<OrcStripeMetadata> readStripesMetadata(
boolean[] globalInc, boolean[] sargColumns) throws IOException {
ArrayList<OrcStripeMetadata> result = new ArrayList<OrcStripeMetadata>(readState.length);
OrcBatchKey stripeKey = new OrcBatchKey(fileId, 0, 0);
for (int stripeIxMod = 0; stripeIxMod < readState.length; ++stripeIxMod) {
stripeKey.stripeIx = stripeIxMod + stripeIxFrom;
OrcStripeMetadata value = metadataCache.getStripeMetadata(stripeKey);
if (value == null || !value.hasAllIndexes(globalInc)) {
counters.incrCounter(Counter.METADATA_CACHE_MISS);
ensureMetadataReader();
StripeInformation si = fileMetadata.getStripes().get(stripeKey.stripeIx);
if (value == null) {
long startTime = counters.startTimeCounter();
value = new OrcStripeMetadata(stripeKey, metadataReader, si, globalInc, sargColumns);
counters.incrTimeCounter(Counter.HDFS_TIME_US, startTime);
value = metadataCache.putStripeMetadata(value);
if (DebugUtils.isTraceOrcEnabled()) {
LlapIoImpl.LOG.info(
"Caching stripe "
+ stripeKey.stripeIx
+ " metadata with includes: "
+ DebugUtils.toString(globalInc));
}
// Create new key object to reuse for gets; we've used the old one to put in cache.
stripeKey = new OrcBatchKey(fileId, 0, 0);
}
// We might have got an old value from cache; recheck it has indexes.
if (!value.hasAllIndexes(globalInc)) {
if (DebugUtils.isTraceOrcEnabled()) {
LlapIoImpl.LOG.info(
"Updating indexes in stripe "
+ stripeKey.stripeIx
+ " metadata for includes: "
+ DebugUtils.toString(globalInc));
}
updateLoadedIndexes(value, si, globalInc, sargColumns);
}
} else {
counters.incrCounter(Counter.METADATA_CACHE_HIT);
}
result.add(value);
consumer.setStripeMetadata(value);
}
return result;
}
int decRef() {
int newRefCount = refCount.decrementAndGet();
if (DebugUtils.isTraceLockingEnabled()) {
LlapIoImpl.LOG.info("Unlocked " + this + "; refcount " + newRefCount);
}
if (newRefCount < 0) {
throw new AssertionError("Unexpected refCount " + newRefCount + ": " + this);
}
return newRefCount;
}
/** Ensures orcReader is initialized for the split. */
private void ensureOrcReader() throws IOException {
if (orcReader != null) return;
Path path = HdfsUtils.getFileIdPath(fs, split.getPath(), fileId);
if (DebugUtils.isTraceOrcEnabled()) {
LOG.info("Creating reader for " + path + " (" + split.getPath() + ")");
}
long startTime = counters.startTimeCounter();
ReaderOptions opts = OrcFile.readerOptions(conf).filesystem(fs).fileMetadata(fileMetadata);
orcReader = EncodedOrcFile.createReader(path, opts);
counters.incrTimeCounter(Counter.HDFS_TIME_US, startTime);
}
/** @return Whether the we can invalidate; false if locked or already evicted. */
@Override
public boolean invalidate() {
while (true) {
int value = refCount.get();
if (value != 0) return false;
if (refCount.compareAndSet(value, EVICTED_REFCOUNT)) break;
}
if (DebugUtils.isTraceLockingEnabled()) {
LlapIoImpl.LOG.info("Invalidated " + this + " due to eviction");
}
return true;
}
int incRef() {
int newRefCount = -1;
while (true) {
int oldRefCount = refCount.get();
if (oldRefCount == EVICTED_REFCOUNT) return -1;
assert oldRefCount >= 0 : "oldRefCount is " + oldRefCount + " " + this;
newRefCount = oldRefCount + 1;
if (refCount.compareAndSet(oldRefCount, newRefCount)) break;
}
if (DebugUtils.isTraceLockingEnabled()) {
LlapIoImpl.LOG.info("Locked " + this + "; new ref count " + newRefCount);
}
return newRefCount;
}
@Override
public DiskRangeList readFileData(DiskRangeList range, long baseOffset, boolean doForceDirect)
throws IOException {
long startTime = counters.startTimeCounter();
DiskRangeList result = orcDataReader.readFileData(range, baseOffset, doForceDirect);
counters.recordHdfsTime(startTime);
if (DebugUtils.isTraceOrcEnabled() && LOG.isInfoEnabled()) {
LOG.info(
"Disk ranges after disk read (file "
+ fileId
+ ", base offset "
+ baseOffset
+ "): "
+ RecordReaderUtils.stringifyDiskRanges(result));
}
return result;
}
@Override
public void returnData(OrcEncodedColumnBatch ecb) {
for (ColumnStreamData[] datas : ecb.getColumnData()) {
if (datas == null) continue;
for (ColumnStreamData data : datas) {
if (data == null || data.decRef() != 0) continue;
if (DebugUtils.isTraceLockingEnabled()) {
for (MemoryBuffer buf : data.getCacheBuffers()) {
LlapIoImpl.LOG.info("Unlocking " + buf + " at the end of processing");
}
}
lowLevelCache.releaseBuffers(data.getCacheBuffers());
CSD_POOL.offer(data);
}
}
// We can offer ECB even with some streams not discarded; reset() will clear the arrays.
ECB_POOL.offer(ecb);
}
/** Determine which stripes to read for a split. Populates stripeIxFrom and readState. */
public void determineStripesToRead() {
// The unit of caching for ORC is (rg x column) (see OrcBatchKey).
List<StripeInformation> stripes = fileMetadata.getStripes();
long offset = split.getStart(), maxOffset = offset + split.getLength();
stripeIxFrom = -1;
int stripeIxTo = -1;
if (LlapIoImpl.LOGL.isDebugEnabled()) {
String tmp = "FileSplit {" + split.getStart() + ", " + split.getLength() + "}; stripes ";
for (StripeInformation stripe : stripes) {
tmp += "{" + stripe.getOffset() + ", " + stripe.getLength() + "}, ";
}
LlapIoImpl.LOG.debug(tmp);
}
int stripeIx = 0;
for (StripeInformation stripe : stripes) {
long stripeStart = stripe.getOffset();
if (offset > stripeStart) {
// We assume splits will never start in the middle of the stripe.
++stripeIx;
continue;
}
if (stripeIxFrom == -1) {
if (DebugUtils.isTraceOrcEnabled()) {
LlapIoImpl.LOG.info(
"Including stripes from " + stripeIx + " (" + stripeStart + " >= " + offset + ")");
}
stripeIxFrom = stripeIx;
}
if (stripeStart >= maxOffset) {
stripeIxTo = stripeIx;
if (DebugUtils.isTraceOrcEnabled()) {
LlapIoImpl.LOG.info(
"Including stripes until "
+ stripeIxTo
+ " ("
+ stripeStart
+ " >= "
+ maxOffset
+ "); "
+ (stripeIxTo - stripeIxFrom)
+ " stripes");
}
break;
}
++stripeIx;
}
if (stripeIxFrom == -1) {
if (LlapIoImpl.LOG.isInfoEnabled()) {
LlapIoImpl.LOG.info("Not including any stripes - empty split");
}
}
if (stripeIxTo == -1 && stripeIxFrom != -1) {
stripeIxTo = stripeIx;
if (DebugUtils.isTraceOrcEnabled()) {
LlapIoImpl.LOG.info(
"Including stripes until "
+ stripeIx
+ " (end of file); "
+ (stripeIxTo - stripeIxFrom)
+ " stripes");
}
}
readState = new boolean[stripeIxTo - stripeIxFrom][][];
}
protected Void performDataRead() throws IOException {
long startTime = counters.startTimeCounter();
if (LlapIoImpl.LOGL.isInfoEnabled()) {
LlapIoImpl.LOG.info("Processing data for " + split.getPath());
}
if (processStop()) {
recordReaderTime(startTime);
return null;
}
counters.setDesc(QueryFragmentCounters.Desc.TABLE, getDbAndTableName(split.getPath()));
orcReader = null;
// 1. Get file metadata from cache, or create the reader and read it.
// Don't cache the filesystem object for now; Tez closes it and FS cache will fix all that
fs = split.getPath().getFileSystem(conf);
fileId = determineFileId(fs, split);
counters.setDesc(QueryFragmentCounters.Desc.FILE, fileId);
try {
fileMetadata = getOrReadFileMetadata();
consumer.setFileMetadata(fileMetadata);
validateFileMetadata();
if (columnIds == null) {
columnIds = createColumnIds(fileMetadata);
}
// 2. Determine which stripes to read based on the split.
determineStripesToRead();
} catch (Throwable t) {
recordReaderTime(startTime);
consumer.setError(t);
return null;
}
if (readState.length == 0) {
consumer.setDone();
recordReaderTime(startTime);
return null; // No data to read.
}
counters.setDesc(QueryFragmentCounters.Desc.STRIPES, stripeIxFrom + "," + readState.length);
// 3. Apply SARG if needed, and otherwise determine what RGs to read.
int stride = fileMetadata.getRowIndexStride();
ArrayList<OrcStripeMetadata> stripeMetadatas = null;
boolean[] globalIncludes = null;
boolean[] sargColumns = null;
try {
globalIncludes = OrcInputFormat.genIncludedColumns(fileMetadata.getTypes(), columnIds, true);
if (sarg != null && stride != 0) {
// TODO: move this to a common method
int[] filterColumns =
RecordReaderImpl.mapSargColumnsToOrcInternalColIdx(sarg.getLeaves(), columnNames, 0);
// included will not be null, row options will fill the array with trues if null
sargColumns = new boolean[globalIncludes.length];
for (int i : filterColumns) {
// filter columns may have -1 as index which could be partition column in SARG.
if (i > 0) {
sargColumns[i] = true;
}
}
// If SARG is present, get relevant stripe metadata from cache or readers.
stripeMetadatas = readStripesMetadata(globalIncludes, sargColumns);
}
// Now, apply SARG if any; w/o sarg, this will just initialize readState.
boolean hasData = determineRgsToRead(globalIncludes, stride, stripeMetadatas);
if (!hasData) {
consumer.setDone();
recordReaderTime(startTime);
return null; // No data to read.
}
} catch (Throwable t) {
cleanupReaders();
consumer.setError(t);
recordReaderTime(startTime);
return null;
}
if (processStop()) {
cleanupReaders();
recordReaderTime(startTime);
return null;
}
// 4. Get data from high-level cache.
// If some cols are fully in cache, this will also give us the modified list of columns to
// read for every stripe (null means read all of them - the usual path). In any case,
// readState will be modified for column x rgs that were fetched from high-level cache.
List<Integer>[] stripeColsToRead = null;
if (cache != null) {
try {
stripeColsToRead = produceDataFromCache(stride);
} catch (Throwable t) {
// produceDataFromCache handles its own cleanup.
consumer.setError(t);
cleanupReaders();
recordReaderTime(startTime);
return null;
}
}
// 5. Create encoded data reader.
// In case if we have high-level cache, we will intercept the data and add it there;
// otherwise just pass the data directly to the consumer.
Consumer<OrcEncodedColumnBatch> dataConsumer = (cache == null) ? this.consumer : this;
try {
ensureOrcReader();
// Reader creating updates HDFS counters, don't do it here.
DataWrapperForOrc dw = new DataWrapperForOrc();
stripeReader = orcReader.encodedReader(fileId, dw, dw, POOL_FACTORY);
stripeReader.setDebugTracing(DebugUtils.isTraceOrcEnabled());
} catch (Throwable t) {
consumer.setError(t);
recordReaderTime(startTime);
cleanupReaders();
return null;
}
// 6. Read data.
// TODO: I/O threadpool could be here - one thread per stripe; for now, linear.
OrcBatchKey stripeKey = new OrcBatchKey(fileId, -1, 0);
for (int stripeIxMod = 0; stripeIxMod < readState.length; ++stripeIxMod) {
if (processStop()) {
cleanupReaders();
recordReaderTime(startTime);
return null;
}
int stripeIx = stripeIxFrom + stripeIxMod;
boolean[][] colRgs = null;
boolean[] stripeIncludes = null;
OrcStripeMetadata stripeMetadata = null;
StripeInformation stripe;
try {
List<Integer> cols = stripeColsToRead == null ? null : stripeColsToRead[stripeIxMod];
if (cols != null && cols.isEmpty()) continue; // No need to read this stripe.
stripe = fileMetadata.getStripes().get(stripeIx);
if (DebugUtils.isTraceOrcEnabled()) {
LlapIoImpl.LOG.info(
"Reading stripe " + stripeIx + ": " + stripe.getOffset() + ", " + stripe.getLength());
}
colRgs = readState[stripeIxMod];
// We assume that NO_RGS value is only set from SARG filter and for all columns;
// intermediate changes for individual columns will unset values in the array.
// Skip this case for 0-column read. We could probably special-case it just like we do
// in EncodedReaderImpl, but for now it's not that important.
if (colRgs.length > 0 && colRgs[0] == SargApplier.READ_NO_RGS) continue;
// 6.1. Determine the columns to read (usually the same as requested).
if (cache == null || cols == null || cols.size() == colRgs.length) {
cols = columnIds;
stripeIncludes = globalIncludes;
} else {
// We are reading subset of the original columns, remove unnecessary bitmasks/etc.
// This will never happen w/o high-level cache.
stripeIncludes = OrcInputFormat.genIncludedColumns(fileMetadata.getTypes(), cols, true);
colRgs = genStripeColRgs(cols, colRgs);
}
// 6.2. Ensure we have stripe metadata. We might have read it before for RG filtering.
boolean isFoundInCache = false;
if (stripeMetadatas != null) {
stripeMetadata = stripeMetadatas.get(stripeIxMod);
} else {
stripeKey.stripeIx = stripeIx;
stripeMetadata = metadataCache.getStripeMetadata(stripeKey);
isFoundInCache = (stripeMetadata != null);
if (!isFoundInCache) {
counters.incrCounter(Counter.METADATA_CACHE_MISS);
ensureMetadataReader();
long startTimeHdfs = counters.startTimeCounter();
stripeMetadata =
new OrcStripeMetadata(
stripeKey, metadataReader, stripe, stripeIncludes, sargColumns);
counters.incrTimeCounter(Counter.HDFS_TIME_US, startTimeHdfs);
stripeMetadata = metadataCache.putStripeMetadata(stripeMetadata);
if (DebugUtils.isTraceOrcEnabled()) {
LlapIoImpl.LOG.info(
"Caching stripe "
+ stripeKey.stripeIx
+ " metadata with includes: "
+ DebugUtils.toString(stripeIncludes));
}
stripeKey = new OrcBatchKey(fileId, -1, 0);
}
consumer.setStripeMetadata(stripeMetadata);
}
if (!stripeMetadata.hasAllIndexes(stripeIncludes)) {
if (DebugUtils.isTraceOrcEnabled()) {
LlapIoImpl.LOG.info(
"Updating indexes in stripe "
+ stripeKey.stripeIx
+ " metadata for includes: "
+ DebugUtils.toString(stripeIncludes));
}
assert isFoundInCache;
counters.incrCounter(Counter.METADATA_CACHE_MISS);
ensureMetadataReader();
updateLoadedIndexes(stripeMetadata, stripe, stripeIncludes, sargColumns);
} else if (isFoundInCache) {
counters.incrCounter(Counter.METADATA_CACHE_HIT);
}
} catch (Throwable t) {
consumer.setError(t);
cleanupReaders();
recordReaderTime(startTime);
return null;
}
if (processStop()) {
cleanupReaders();
recordReaderTime(startTime);
return null;
}
// 6.3. Finally, hand off to the stripe reader to produce the data.
// This is a sync call that will feed data to the consumer.
try {
// TODO: readEncodedColumns is not supposed to throw; errors should be propagated thru
// consumer. It is potentially holding locked buffers, and must perform its own cleanup.
// Also, currently readEncodedColumns is not stoppable. The consumer will discard the
// data it receives for one stripe. We could probably interrupt it, if it checked that.
stripeReader.readEncodedColumns(
stripeIx,
stripe,
stripeMetadata.getRowIndexes(),
stripeMetadata.getEncodings(),
stripeMetadata.getStreams(),
stripeIncludes,
colRgs,
dataConsumer);
} catch (Throwable t) {
consumer.setError(t);
cleanupReaders();
recordReaderTime(startTime);
return null;
}
}
// Done with all the things.
recordReaderTime(startTime);
dataConsumer.setDone();
if (DebugUtils.isTraceMttEnabled()) {
LlapIoImpl.LOG.info("done processing " + split);
}
// Close the stripe reader, we are done reading.
cleanupReaders();
return null;
}