Add a memory bound FileStatisticsCache for the Listing Table#20047
Open
mkleen wants to merge 49 commits intoapache:mainfrom
Open
Add a memory bound FileStatisticsCache for the Listing Table#20047mkleen wants to merge 49 commits intoapache:mainfrom
mkleen wants to merge 49 commits intoapache:mainfrom
Conversation
github-actions
bot
added
documentation
mkleen
force-pushed
the
file-stats-cache
branch
from
a66420a to
3b33739
Compare
mkleen
force-pushed
the
file-stats-cache
branch
from
3b33739 to
8e5560b
Compare
github-actions
bot
removed
the
documentation
mkleen
force-pushed
the
file-stats-cache
branch
2 times, most recently
from
e273afc to
b297378
Compare
github-actions
bot
added
the
documentation
mkleen
changed the title
mkleen
changed the title
mkleen
changed the title
mkleen
changed the title
mkleen
changed the title
kosiew
requested changes
Feb 4, 2026
mkleen
force-pushed
the
file-stats-cache
branch
from
59c6bce to
4542db8
Compare
Contributor
Author
|
@kosiew Thank you for the feedback! |
Contributor
Author
|
@kosiew Anything else needed to get this merged? Another approval maybe? |
mkleen
force-pushed
the
file-stats-cache
branch
from
205f96c to
92899a7
Compare
martin-g
reviewed
Feb 10, 2026
| impl<T: DFHeapSize> DFHeapSize for Arc<T> { | ||
| fn heap_size(&self) -> usize { | ||
| // Arc stores weak and strong counts on the heap alongside an instance of T | ||
| 2 * size_of::<usize>() + size_of::<T>() + self.as_ref().heap_size() |
Member
There was a problem hiding this comment.
This won't be accurate.
let a1 = Arc::new(vec![1, 2, 3]);
let a2 = a1.clone();
let a3 = a1.clone();
let a4 = a3.clone();
// this should be true because all `a`s point to the same object in memory
// but the current implementation does not detect this and counts them separately
assert_eq!(a4.heap_size(), a1.heap_size() + a2.heap_size() + a3.heap_size() + a4.heap_size());The only solution I imagine is the caller to keep track of the pointer addresses which have been "sized" and ignore any Arc's which point to an address which has been "sized" earlier.
Contributor
Author
There was a problem hiding this comment.
Good catch! I took this implementation from https://github.com/apache/arrow-rs/blob/main/parquet/src/file/metadata/memory.rs#L97-L102 . I would suggest to also do a follow-up here. We are planing anyway to restructure the whole heap size estimation.
Contributor
Author
|
@martin-g Thanks for this great review! I am on it. |
mkleen
force-pushed
the
file-stats-cache
branch
from
92899a7 to
2e3aff9
Compare
mkleen
force-pushed
the
file-stats-cache
branch
from
365cb2f to
9b48bd8
Compare
Contributor
Author
|
@adriangb Could please do another benchmark run? I have no regressions anymore running |
Contributor
|
run benchmark clickbench_partitioned |
Contributor
Author
@Dandandan Thank you, i broke the linter. It's fixed now. Does this affect the benchmark run? |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing file-stats-cache (3b64005) to 244f891 (merge-base) diff using: clickbench_partitioned File an issue against this benchmark runner |
|
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)Details
Resource Usageclickbench_partitioned — base (merge-base)
clickbench_partitioned — branch
File an issue against this benchmark runner |
Contributor
Author
|
This looks much better now: There are a few minor regressions which could be just noise. Could we do a benchmark second run? |
Contributor
Author
|
@kosiew I added all the changes you requested. Could you do another review please? |
Contributor
|
run benchmark clickbench_partitioned |
|
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing file-stats-cache (3b64005) to 244f891 (merge-base) diff using: clickbench_partitioned File an issue against this benchmark runner |
Contributor
|
@mkleen you should be able to trigger benchmark runs now fwiw |
|
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)Details
Resource Usageclickbench_partitioned — base (merge-base)
clickbench_partitioned — branch
File an issue against this benchmark runner |
Contributor
Author
|
Ok, the benchmarks are looking good now again. At the moment we have 20 mb default cache size, with 10 mb the stats for clickhouse_partitioned did not fit yet. I wonder what would be a reasonable default value? |
kosiew
requested changes
Apr 16, 2026
Contributor
kosiew
left a comment
kosiew
left a comment
There was a problem hiding this comment.
Thanks for the updates here, this is shaping up nicely.
I still have a couple of concerns.
Some previously noted issues appear to be intentionally deferred, like the Arc heap size double-counting and the empty string parsing case. Could you clarify whether those are planned for a follow-up PR?
I also ran into two additional issues while reviewing the final diff, including what looks like a functional regression in the reader path cache wiring. Because of that, I am not yet confident that the branch is free of new problems.
Would love your thoughts on the points below.
| let table = ListingTable::try_new(config)?.with_definition(sql_definition); | ||
| let table = ListingTable::try_new(config)? | ||
| .with_definition(sql_definition) | ||
| .with_cache(self.runtime_env().cache_manager.get_file_statistic_cache()); |
Contributor
There was a problem hiding this comment.
I noticed that only the registered-listing-table path was updated to call .with_cache(...). Was it intentional that the generic reader path in _read_type() still uses ListingTable::try_new(config)? without attaching the runtime cache?
With ListingTable now defaulting collected_statistics to None, could this mean that SessionContext::read_parquet, read_csv, read_json, read_avro, and similar methods no longer benefit from the file statistics cache?
If so, would this be considered a regression in the default-cache rollout? How do you see this affecting users who rely on those reader paths?
Contributor
Author
There was a problem hiding this comment.
No this was not intentional. This needs to be fixed. Thanks a lot!
| Some("50M".to_owned()), | ||
| Some("1M".to_owned()), | ||
| None, | ||
| Some("1M".to_owned()), |
Contributor
There was a problem hiding this comment.
I see that DEFAULT_FILE_STATISTICS_MEMORY_LIMIT was increased to 20 MiB, but RuntimeEnvBuilder::entries() still hard-codes Some("1M".to_owned()) for datafusion.runtime.file_statistics_cache_limit.
Would it make sense to update entries() to match the new default?
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
8 participants
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.
Which issue does this PR close?
This change introduces a default
FileStatisticsCacheimplementation for the Listing-Table with a size limit, implementing the following steps following #19052 (comment) :Add heap size estimation for file statistics and the relevant data types used in caching (This is temporary until Add a crate for HeapSize trait arrow-rs#9138 is resolved)
Redesign
DefaultFileStatisticsCacheto use aLruQueueto make it memory-bound following Adds memory-bound DefaultListFilesCache #18855Introduce a size limit and use it together with the heap-size to limit the memory usage of the cache
Move
FileStatisticsCachecreation intoCacheManager, making it session-scoped and shared across statements and tables.Disable caching in some of the SQL-logic tests where the change altered the output result, because the cache is now session-scoped and not query-scoped anymore.
Closes Add a default
FileStatisticsCacheimplementation for theListingTable#19217Closes Add limit to
DefaultFileStatisticsCache#19052Rationale for this change
See above.
What changes are included in this PR?
See above.
Are these changes tested?
Yes.
Are there any user-facing changes?
A new runtime setting
datafusion.runtime.file_statistics.cache_limit