[SPARK-9419] ShuffleMemoryManager and MemoryStore should track memory on a per-task, not per-thread, basis #7734
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
… cleanup is performed
|
Note that this still needs test cases. I also need to audit PythonRDD, RRDD, ScriptTransformation, and PipedRDD to make sure that they propagate their parent threads' TaskContext thread locals. |
|
/cc @andrewor14 for first-impressions on the approach here; just looking for high-level feedback on whether this is the right route to fix some of these problems of memory being freed in different threads than it was allocated in. |
17 tasks
|
Test build #38762 has finished for PR 7734 at commit
|
|
Test build #38766 has finished for PR 7734 at commit
|
|
@lianhuiwang, you might be interested in this change given the related work that you're doing at #7130. |
| @@ -263,11 +263,6 @@ private[spark] class PythonRDD( | |||
| if (!worker.isClosed) { | |||
| Utils.tryLog(worker.shutdownOutput()) | |||
| } | |||
| } finally { | |||
There was a problem hiding this comment.
/cc @davies for this PySpark change.
|
LGTM |
|
As discussed offline, one thing that's missing is to propagate the task context to other threads. |
|
Test build #38792 has finished for PR 7734 at commit
|
|
I've merged this. Thanks! |
|
Just for future readers of this PR: note that the ScriptTransformation case was already handled in a previous PR, which is why there's not a change for that here. |
asfgit
pushed a commit
that referenced
this pull request
Jul 30, 2015
This pull request enables Unsafe mode by default in Spark SQL. In order to do this, we had to fix a number of small issues: **List of fixed blockers**: - [x] Make some default buffer sizes configurable so that HiveCompatibilitySuite can run properly (#7741). - [x] Memory leak on grouped aggregation of empty input (fixed by #7560 to fix this) - [x] Update planner to also check whether codegen is enabled before planning unsafe operators. - [x] Investigate failing HiveThriftBinaryServerSuite test. This turns out to be caused by a ClassCastException that occurs when Exchange tries to apply an interpreted RowOrdering to an UnsafeRow when range partitioning an RDD. This could be fixed by #7408, but a shorter-term fix is to just skip the Unsafe exchange path when RangePartitioner is used. - [x] Memory leak exceptions masking exceptions that actually caused tasks to fail (will be fixed by #7603). - [x] ~~https://issues.apache.org/jira/browse/SPARK-9162, to implement code generation for ScalaUDF. This is necessary for `UDFSuite` to pass. For now, I've just ignored this test in order to try to find other problems while we wait for a fix.~~ This is no longer necessary as of #7682. - [x] Memory leaks from Limit after UnsafeExternalSort cause the memory leak detector to fail tests. This is a huge problem in the HiveCompatibilitySuite (fixed by f4ac642a4e5b2a7931c5e04e086bb10e263b1db6). - [x] Tests in `AggregationQuerySuite` are failing due to NaN-handling issues in UnsafeRow, which were fixed in #7736. - [x] `org.apache.spark.sql.ColumnExpressionSuite.rand` needs to be updated so that the planner check also matches `TungstenProject`. - [x] After having lowered the buffer sizes to 4MB so that most of HiveCompatibilitySuite runs: - [x] Wrong answer in `join_1to1` (fixed by #7680) - [x] Wrong answer in `join_nulls` (fixed by #7680) - [x] Managed memory OOM / leak in `lateral_view` - [x] Seems to hang indefinitely in `partcols1`. This might be a deadlock in script transformation or a bug in error-handling code? The hang was fixed by #7710. - [x] Error while freeing memory in `partcols1`: will be fixed by #7734. - [x] After fixing the `partcols1` hang, it appears that a number of later tests have issues as well. - [x] Fix thread-safety bug in codegen fallback expression evaluation (#7759). Author: Josh Rosen <joshrosen@databricks.com> Closes #7564 from JoshRosen/unsafe-by-default and squashes the following commits: 83c0c56 [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-by-default f4cc859 [Josh Rosen] Merge remote-tracking branch 'origin/master' into unsafe-by-default 963f567 [Josh Rosen] Reduce buffer size for R tests d6986de [Josh Rosen] Lower page size in PySpark tests 013b9da [Josh Rosen] Also match TungstenProject in checkNumProjects 5d0b2d3 [Josh Rosen] Add task completion callback to avoid leak in limit after sort ea250da [Josh Rosen] Disable unsafe Exchange path when RangePartitioning is used 715517b [Josh Rosen] Enable Unsafe by default
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Spark's ShuffleMemoryManager and MemoryStore track memory on a per-thread basis, which causes problems in the handful of cases where we have tasks that use multiple threads. In PythonRDD, RRDD, ScriptTransformation, and PipedRDD we consume the input iterator in a separate thread in order to write it to an external process. As a result, these RDD's input iterators are consumed in a different thread than the thread that created them, which can cause problems in our memory allocation tracking. For example, if allocations are performed in one thread but deallocations are performed in a separate thread then memory may be leaked or we may get errors complaining that more memory was allocated than was freed.
I think that the right way to fix this is to change our accounting to be performed on a per-task instead of per-thread basis. Note that the current per-thread tracking has caused problems in the past; SPARK-3731 (#2668) fixes a memory leak in PythonRDD that was caused by this issue (that fix is no longer necessary as of this patch).