core: Provide mechanism to cache manifest file content

[rizaon]

This is a draft PR for ManifestCache implementation.

The ManifestCache interface closely follow com.github.benmanes.caffeine.cache.Cache interface, with addition of specifying maxContentLength(). If stream length is longer than maxContentLength(), AvroIO will skip caching it to avoid memory pressure. An example of implementation, CaffeineManifestCache, can be seen in TestAvroFileSplit.java.

I will tidy this up and add documentations as we go. Looking forward for any feedback.

Closes #4508

[rizaon]

485e2ab Switch ManifestCache class to use Guava Cache instead. This is easier to do in iceberg-api since we can simply include the Guava cache classes into the shadow jar of iceberg-bundled-guava.

[rdblue]

Switching to use more Guava classes is probably not a good idea.

@rizaon, do you have use cases where this has helped? If so, what were they? This adds quite a bit of complexity and memory overhead that we purposely avoided up to now so I want to make sure it is worth the change.

Have you considered adding a caching FileIO instance? That could be used for more use cases than just manifest files and job planning. For example, a FileIO read-through cache could cache delete files that might be reused for tasks. This could be configured by file path, making it easy to determine what to cache. Plus, we could use more options than just in-memory, like a distributed cache or local disk.

[rizaon]

Hi @rdblue, thank you for your feedback.

We found a slow query compilation issue against the Iceberg table in our recent Apache Impala build. Impala uses Iceberg's HiveCatalog and HadoopFileIO instance with an S3A input stream to access data from S3. We did a full 10 TB TPC-DS benchmark and found that query compilation can go for several seconds, while it used to be less than a second with native hive tables. This slowness in single query compilation is due to the requirement to call planFiles several times, even for scan nodes targetting the same table. We also see several socket read operations that spend hundreds of milliseconds during planFiles, presumably due to S3A HTTP HEAD request overhead and backward seek overhead (issue #4508). This is especially hurt fast-running queries.

We tried this caching solution and it help speed up Impala query compilation almost 5x faster compared to without on Iceberg tables. Our original solution, however, is to put a Caffeine cache as a singleton in AvroIO.java. I thought it is better to supply the cache from outside.

I have not considered the solution of adding a caching FileIO instance. I'm pretty new to the Iceberg codebase but interested to follow up on that if it can yield a better integration. Will it require a new class of Catalog/Table as well, or can we improve on the existing HiveCatalog & HadoopFileIO?

[rizaon]

A relevant Impala's JIRA is here:
https://issues.apache.org/jira/browse/IMPALA-11171

[rdblue]

@rizaon, caching in the FileIO layer would be much more general. You could do things like detect that the file size is less than some threshold and cache it in memory, or detect file names under metadata and cache those. I highly recommend looking into the FileIO API!

[rizaon]

c032b7a implement caching as a new FileIO class, CachingHadoopFileIO. A new Tables class, CachingHadoopTables, is also added to assist with testing.

We tried to avoid lazyStats() call as much as possible by checking the cache first before checking for stream length. This comes with a risk of keeping stale data in memory when the actual file is already deleted, as shown in CachingHiveTableTest.testMissingMetadataWontCauseHang(). This is probably fine, given the immutable nature of iceberg metadata. Otherwise, application should invalidate cache entry first to force re-read / lazyStats() call.

[rizaon]

Hello @rdblue. This PR is ready for review.

Please let me know if there is any new feedback or request. I'm happy to follow up.

[rdblue]

Is there a way to pass in the cache rather than making it static?

@jackye1995, do you have any ideas here?

[rizaon]

This remain static in 47b8008. Please let me know if there is a better way to access the cache from Catalog object or outside.

[rdblue]

@rizaon, there are a couple of PRs that should help you with this. #4608 adds IOUtil.readFully and IOUtil.readRemaining that you can use to fill buffers from an input stream. I've also opened #4623 to add the ByteBufferInputStream classes I was talking about. Those should allow you to easily split the data across several small allocations and wrap them in an InputStream implementation.

[rizaon]

Got it, thank you. Will rebase and update this PR once they are merged.

[rizaon]

In the last github check runs, there was a checkstye issue and exception handling when given FileIO does not implement properties() method.

I have rebased this PR and add ba7329d to fix this issue.

[danielcweeks]

@rizaon Thanks for your contribution here and keeping this updated! I'm excited to see how this works out.

[rizaon]

@danielcweeks thank you for accepting this PR!