Spark lists all leaf node even in partitioned data
As soon as spark is given a directory to read from it issues call to listLeafFiles (org/apache/spark/sql/execution/datasources/fileSourceInterfaces.scala). This in turn calls fs.listStatus which makes an api call to get list of files and directories. Now for each directory this method is called again. This hapens recursively until no directories are left. This by design works good in a HDFS system. But works bad in s3 since list file is an RPC call. S3 on other had supports get all files by prefix, which is exactly what we need.
So for example if we had above directory structure with 1 year worth of data with each directory for hour and 10 sub directory we would have , 365 * 24 * 10 = 87k api calls, this can be reduced to 138 api calls given that there are only 137000 files. Each s3 api calls return 1000 files.
Code:
org/apache/hadoop/fs/s3a/S3AFileSystem.java
public FileStatus[] listStatusRecursively(Path f) throws FileNotFoundException,
IOException {
String key = pathToKey(f);
if (LOG.isDebugEnabled()) {
LOG.debug("List status for path: " + f);
}
final List<FileStatus> result = new ArrayList<FileStatus>();
final FileStatus fileStatus = getFileStatus(f);
if (fileStatus.isDirectory()) {
if (!key.isEmpty()) {
key = key + "/";
}
ListObjectsRequest request = new ListObjectsRequest();
request.setBucketName(bucket);
request.setPrefix(key);
request.setMaxKeys(maxKeys);
if (LOG.isDebugEnabled()) {
LOG.debug("listStatus: doing listObjects for directory " + key);
}
ObjectListing objects = s3.listObjects(request);
statistics.incrementReadOps(1);
while (true) {
for (S3ObjectSummary summary : objects.getObjectSummaries()) {
Path keyPath = keyToPath(summary.getKey()).makeQualified(uri, workingDir);
// Skip over keys that are ourselves and old S3N _$folder$ files
if (keyPath.equals(f) || summary.getKey().endsWith(S3N_FOLDER_SUFFIX)) {
if (LOG.isDebugEnabled()) {
LOG.debug("Ignoring: " + keyPath);
}
continue;
}
if (objectRepresentsDirectory(summary.getKey(), summary.getSize())) {
result.add(new S3AFileStatus(true, true, keyPath));
if (LOG.isDebugEnabled()) {
LOG.debug("Adding: fd: " + keyPath);
}
} else {
result.add(new S3AFileStatus(summary.getSize(),
dateToLong(summary.getLastModified()), keyPath,
getDefaultBlockSize(f.makeQualified(uri, workingDir))));
if (LOG.isDebugEnabled()) {
LOG.debug("Adding: fi: " + keyPath);
}
}
}
for (String prefix : objects.getCommonPrefixes()) {
Path keyPath = keyToPath(prefix).makeQualified(uri, workingDir);
if (keyPath.equals(f)) {
continue;
}
result.add(new S3AFileStatus(true, false, keyPath));
if (LOG.isDebugEnabled()) {
LOG.debug("Adding: rd: " + keyPath);
}
}
if (objects.isTruncated()) {
if (LOG.isDebugEnabled()) {
LOG.debug("listStatus: list truncated - getting next batch");
}
objects = s3.listNextBatchOfObjects(objects);
statistics.incrementReadOps(1);
} else {
break;
}
}
} else {
if (LOG.isDebugEnabled()) {
LOG.debug("Adding: rd (not a dir): " + f);
}
result.add(fileStatus);
}
return result.toArray(new FileStatus[result.size()]);
}
/org/apache/spark/sql/execution/datasources/fileSourceInterfaces.scala
def listLeafFiles(fs: FileSystem, status: FileStatus, filter: PathFilter): Array[FileStatus] = {
logTrace(s"Listing ${status.getPath}")
val name = status.getPath.getName.toLowerCase
if (shouldFilterOut(name)) {
Array.empty[FileStatus]
}
else {
val statuses = {
val stats = if(fs.isInstanceOf[S3AFileSystem]){
logWarning("Using Monkey patched version of list status")
println("Using Monkey patched version of list status")
val a = fs.asInstanceOf[S3AFileSystem].listStatusRecursively(status.getPath)
a
// Array.empty[FileStatus]
}
else{
val (dirs, files) = fs.listStatus(status.getPath).partition(_.isDirectory)
files ++ dirs.flatMap(dir => listLeafFiles(fs, dir, filter))
}
if (filter != null) stats.filter(f => filter.accept(f.getPath)) else stats
}
// statuses do not have any dirs.
statuses.filterNot(status => shouldFilterOut(status.getPath.getName)).map {
case f: LocatedFileStatus => f
// NOTE:
//
// - Although S3/S3A/S3N file system can be quite slow for remote file metadata
// operations, calling `getFileBlockLocations` does no harm here since these file system
// implementations don't actually issue RPC for this method.
//
// - Here we are calling `getFileBlockLocations` in a sequential manner, but it should not
// be a big deal since we always use to `listLeafFilesInParallel` when the number of
// paths exceeds threshold.
case f => createLocatedFileStatus(f, fs.getFileBlockLocations(f, 0, f.getLen))
}
}
}
To clarify Gaurav's answer, that code snipped is from Hadoop branch-2, Probably not going to surface until Hadoop 2.9 (see HADOOP-13208); and someone needs to update Spark to use that feature (which won't harm code using HDFS, just won't show any speedup there).
One thing to consider is: what makes a good file layout for Object Stores.
- Don't have deep directory trees with only a few files per directory
- Do have shallow trees with many files
- Consider using the first few characters of a file for the most changing value (such as day/hour), rather than the last. Why? Some object stores appear to use the leading characters for their hashing, not the trailing ones ... if you give your names more uniqueness then they get spread out over more servers, with better bandwidth/less risk of throttling.
- If you are using the Hadoop 2.7 libraries, switch to s3a:// over s3n://. It's already faster, and getting better every week, at least in the ASF source tree.
Finally, Apache Hadoop, Apache Spark and related projects are all open source. Contributions are welcome. That's not just the code, it's documentation, testing, and, for this performance stuff, testing against your actual datasets. Even giving us details about what causes problems (and your dataset layouts) is interesting.