...

This page is meant as a template for writing a KIP. To create a KIP choose Tools->Copy on this page and modify with your content and replace the heading with the next KIP number and a description of your issue. Replace anything in italics with your own description.

Status

Current state: "DraftIn Discussion"

Discussion thread: : https://lists.apache.org/thread/jfj4fn09v5d1n9s7vpt1twht9jodfvlkhere [Change the link from the KIP proposal email archive to your own email thread]

JIRA:

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

Motivation

in In Tiered Storage, indexes are uploaded together with the segment log (on the same operation RemoteStorageManager#copyLogSegment ) when segments become available for uploading to the remote tier.

However, when consumer fetching hits remote data, segments then each segment index is fetched individually and cached in - memory at the broker. Even though fetching segment indexes is an action that happens at least once per segment read, the fact that each segment index is fetched independently can lead to an increased latency when fetching across individually can increase latency as fetching cross segments. This can become a more pressing issue when the segment sizes are smaller than the default (e.g. 10s of MBs).

Remote Storage Manager implementations (referred to as pluginsPlugins) can optimize how segment indexes are uploaded given that all indexes are uploaded on the same operation, and are immutable. For  For instance, plugins can choose to upload the segment indexes as part of the same object and fetch them by range.

HoweverNevertheless, fetching segment indexes individually is more expensive given that remote fetch round-trip is comparatively the most expensive operation in terms of latency and carries a (usually small, but additional) cost.  

Public Interfaces

Briefly list any new interfaces that will be introduced as part of this proposal or any existing interfaces that will be removed or changed. The purpose of this section is to concisely call out the public contract that will come along with this feature.

A public interface is any change to the following:

• Binary log format

• The network protocol and api behavior

• Any class in the public packages under clientsConfiguration, especially client configuration

  • org/apache/kafka/common/serialization

  • org/apache/kafka/common

  • org/apache/kafka/common/errors

  • org/apache/kafka/clients/producer

  • org/apache/kafka/clients/consumer (eventually, once stable)

• Monitoring

• Command line tools and arguments

• Anything else that will likely break existing users in some way when they upgrade

Proposed Changes

To reduce the impact on latency and the number of calls to remote objects, this KIP proposes to include new operations to the RemoteStorageManager  API to fetch indexes at once allowing plugin implementations to optimize the number of objects, number of requests and latency.

Public Interfaces

Add the following method to RemoteStorageManager API:

    /**
     * Returns a set of indexes available for a log segment.
     * If the index type is not available (e.g. transaction index), the resulting map will not include it.
     */
    default Map<IndexType, InputStream> fetchIndexes(RemoteLogSegmentMetadata remoteLogSegmentMetadata, Set<IndexType> indexTypes) throws RemoteStorageException {
        Map<IndexType, InputStream> indexes = new HashMap<>(indexTypes.size());
        for (IndexType indexType: indexTypes) {
            try {
                indexes.put(indexType, fetchIndex(remoteLogSegmentMetadata, indexType));
            } catch (RemoteResourceNotFoundException e) {
                // log
            }
        }
        return indexes;
    }

    
    Set<IndexType> INDEX_TYPES = Arrays.stream(IndexType.values()).collect(Collectors.toSet());

    /**
     * Returns all indexes available for a log segment
     * If the index type is not available (e.g. transaction index), the resulting map will not include it.
     */
    Map<IndexType, InputStream> fetchAllIndexes(RemoteLogSegmentMetadata remoteLogSegmentMetadata) throws RemoteStorageException {
        return fetchIndexes(remoteLogSegmentMetadata, INDEX_TYPES);
    }

Both APIs are proposed for current and future tooling and versions of the Tiered Storage framework optimize how indexes are fetched.

Proposed Changes

Apart from including the new APIs, this proposal includes refactoring RemoteIndexCache to work with the new API:

The remote index cache has an in-memory map between Segment IDs and a set of indexes (i.e. Entry) containing Offset, Timestamp, and Transaction Index if exists. The cache is backed by files used to maintain the cache across restarts.

Instead of fetching each individual index whenever a file is not found locally, if the entry is invalid (e.g. no files found, not all files found, corrupted files) then the proposed operation to fetch a set of index types is used to get all indexes at once, persist them, and load the cacheDescribe the new thing you want to do in appropriate detail. This may be fairly extensive and have large subsections of its own. Or it may be a few sentences. Use judgement based on the scope of the change.

Compatibility, Deprecation, and Migration Plan

• What impact (if any) will there be on existing users?
• If we are changing behavior how will we phase out the older behavior?
• If we need special migration tools, describe them here.
• When will we remove the existing behavior?

Test Plan

To be backward compatible, a default implementation mimicking the current behavior is provided.

Test Plan

The new methods will have to be implemented and tested by the RSM implementations. As RemoteIndexCache  implementation will be adapted to the new APIs, the current set of tests should serve as a validation that functionality isn't broken.

Tiered Storage integration tests will be extended to cover the new APIs and test it similarly to the existing ones.Describe in few sentences how the KIP will be tested. We are mostly interested in system tests (since unit-tests are specific to implementation details). How will we know that the implementation works as expected? How will we know nothing broke?

Rejected Alternatives

If there are alternative ways of accomplishing the same thing, what were they? The purpose of this section is to motivate why the design is the way it is and not some other way.

Cache Indexes at the RSM Plugin level

RSM implementations could cache indexes at the plugin level, so if the next request is for another index of the same segment, it can be served from the cache.

Although this solves the issue with multiple requests to the same object, it duplicates the effort of keeping 2 caches for indexes, one at the broker level, and one at the plugin level.

As currently this is managed by the broker, and unless this responsibility is delegated (if needed) to the plugin, then the proposed approach stands.