Skip to content

Latest commit

 

History

History
140 lines (98 loc) · 6.38 KB

listing-performance.md

File metadata and controls

140 lines (98 loc) · 6.38 KB
Raw

Maximing Hadoop API List performance (especially on object stores)

Summary

  • Embrace the incremental list operations for better scale and support of paged downloads (object storage), adopting

  • Try to do useful work in each iteration, to make the most of of clients which do asynchronous fetch of the next page of results.

  • Be prepared for listing failures (AccessDenied, FileNotFound) to surface in that first hasNext()/next() call; as asynchronous IO may delay the problem being observed in the list operation.

  • RemoteIterator<T> implementations which implement Closeable MUST be closed. That is as important for local file:// filesystems as object stores.

  • RemoteIterator<T> implementations which implement statistics collection can be queried for them on builds with the HADOOP-16830 API, and will print them in their toString() calls. Consider logging at debug to see those values.

Avoid

  • Discarding information and asking the FS for it later in getFileStatus() calls
  • globStatus()
  • Probing for directories existing (getFileStatus(), exists(), isDirectory()) before executing the list call. Do the list, and expect failures to surface in the list call.

Embrace

Use in order of preference

  • listFiles(path, recursive = true) -deep listing, avoids any simulation of directories. S3A: O(files) with async page fetch irrespective of directory structure; elsewhere incremental treewalk (which could do async work in future)

  • listFiles(path, recursive = false) -shallow listing, potentially efficient.

  • listStatusIterator(path) - a direct replacement for FileSystem.listStatus(): FileStatus[]

And for when you want store location: listLocatedStatus().

Mapreduce LocatedFileStatusFetcher can retrieve files in parallel to build its list up (and will provide IOStatistics too). (org.apache.hadoop.mapred.LocatedFileStatusFetcher) It is/will be tagged as @Public/@Unstable, because applications/librareies turn out to have been using it (hello Parquet!). It is/will be an IOStatisticsSource too, providing stats on the list operations, though sadly not the globStatus() phase.

How to Parallelise

How best to parallelise your own listing code, if you are doing it to treewalk

  • Use an incremental list operation.
  • Have a (limited) set of worker threads doing their own processing/scanning, reading from a blocking queue and feeding the results into another queue
  • Feed the listing entries into the queue of child directories to scan.
  • Provide another RemoteIterator<> to access the results, reading from the result queue.
  • Which should implement Closeable.close() if you want to allow faster abort of sequence from the reader.
  • Otherwise: Add a way to stop the scan process.

Application code to :

  • Go through the remote iterator, being prepared for failures to not surface until the first
  • Do something with each entry, the better to make use of asynchronous fetching.
  • Consider measuring list performance. Take a look at org.apache.hadoop.util.DurationInfo if you want something minimal.

Making the best use of RemoteIterator<> listings

  1. Consider that the read may block on the initial list call, the hasNext() or the next() call.
  2. There's some ambiguity as to when FileNotFoundException and similar events may surface(*). It may be in the actual evaluation where FileNotFoundException surfaces, because it was only in an asynchronous listing that the operation failed.

Some iterators may be Closeable for a faster/cleaner shutdown of background IO. Consider checking to see if they do, Closeable, cast and close, e,g through IOUtils.cleanupWithLogger(LOGGER, (Closeable) it);.

This isn't just for the object stores; once we hook the iterable java.nio.Files.newDirectoryStream(path) listing to listStatusIterator(), we will need to close DirectoryStream at the end to avoid leaking references. The FS APIs can do all the passthrough, but the callers will need to wrap up.

RemoteIterators and IOStatistics

Once HADOOP-16380 is in, some RemoteIterators willimplement IOStatisticSource, and, if invoked, may return either null or an IOStatistics implementation.

Specifically:

  • The S3A listing operations will report on the number of http list/list-continuation requests made, and the mix/mean/max values.
  • The mapping/transforming wrapper iterators in org.apache.hadoop.util.functional all implement IOStatisticsSource and will pass through the getIOStatistics() call. They will also forward Closeable.close().

You can use IOStatisticsLogging to log these, or collect results into an IOStatisticsSnapshot for aggregation and reporting. his snapshot is Serializable as java or JSON, so total statisics can be reported back.

Without using the new API, consider logging the iterator's .toString() value at debug-level, if you want to get performance statistics from them.

Implementors notes

General

  • Focus on listStatus() and listLocatedStatus().

Asynchronous listing

  • Use a filesystem thread pool Rather than spin up threads of your own.
  • Try to prefetch as much as you can.
  • But be aware some code probes for a dir having children with a listStatus(path).hasNext() operation, so that initial scan may be something to make fast.
  • Do not rely on close() being invoked.

IOStatistics

  • Collect statistics, report them through IOStatistics and toString()
  • Especially for blocking/async IO operations against remote stores. Identifying listing-related latencies is invaluable for performance tuning query planning
  • Wrap with o.a.h.fs.functional.RemoteStatistics wrappers, which pass through IOStatistics and Closeable APIs. You should be able to use these for most of your simple 1:1 map transformations — unless you need to do things with exception handling/transformation.