[
https://issues.apache.org/jira/browse/HBASE-30406?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Work on HBASE-30406 started by Vladimir Rodionov.
-------------------------------------------------
> Add cache access notification support to CacheEngine and cache topologies
> -------------------------------------------------------------------------
>
> Key: HBASE-30406
> URL: https://issues.apache.org/jira/browse/HBASE-30406
> Project: HBase
> Issue Type: New Feature
> Components: BlockCache
> Reporter: Vladimir Rodionov
> Assignee: Vladimir Rodionov
> Priority: Major
> Fix For: 4.0.0-alpha-1
>
>
> h2. Description
> {{CacheEngine}} currently provides {{getBlock()}} for retrieving cached
> blocks, but there is no lightweight mechanism to notify a cache engine that a
> block has been accessed without retrieving the block from that engine.
> This is required for cache topologies where the same block may reside in
> multiple tiers.
> In particular, {{TieredInclusiveTopology}} keeps a block in both L1 and L2.
> When a request is satisfied from L1, L2 does not observe the access. As a
> result, the L2 cache engine cannot update access-based eviction, recency,
> frequency, or admission metadata for the block.
> Add a {{touch(BlockCacheKey)}} operation to {{CacheEngine}}. The operation
> provides an access notification without retrieving the cached object and
> without affecting normal cache hit/miss accounting.
> Access propagation should be topology-specific. {{CacheTopology}} should
> provide a hook for handling successful cache accesses.
> {{TopologyBackedCacheAccessService}} should notify the topology when a block
> is found, allowing the topology to propagate the access to other tiers when
> required.
> For {{TieredInclusiveTopology}}, an L1 hit should result in a {{touch()}}
> notification to L2. No cross-tier touch is required for
> {{TieredExclusiveTopology}} or {{SingleTierTopology}}.
> h2. Proposed Changes
> * Add {{CacheEngine.touch(BlockCacheKey)}} with a default no-op
> implementation.
> * Add a topology-level access notification hook, e.g.
> {{CacheTopology.handleAccess(...)}}.
> * Update {{TopologyBackedCacheAccessService}} to notify the topology after a
> successful cache lookup.
> * Update {{TieredInclusiveTopology}} to propagate an L1 hit to L2 using
> {{touch()}}.
> * Implement {{touch()}} for cache engines that maintain access-based metadata.
> * Add tests for inclusive, exclusive, and single-tier topology behavior.
> h2. Semantics
> * {{touch()}} is a lightweight access notification and does not retrieve the
> cached object.
> * {{touch()}} is best-effort. If the block is no longer present in the target
> engine, the operation has no effect.
> * {{touch()}} must not increment normal cache hit/miss statistics.
> * An L1 hit in {{TieredInclusiveTopology}} causes L2 to receive {{touch()}}.
> * An L2 hit does not require an additional {{touch()}} because L2 already
> observes the access through {{getBlock()}}.
> * {{TieredExclusiveTopology}} does not propagate L1 hits to L2 because the
> block is not expected to be present in both tiers.
> * {{SingleTierTopology}} requires no additional access propagation.
> * Cache engines that do not need access notifications retain the default
> no-op behavior.
> h2. Acceptance Criteria
> * {{CacheEngine}} exposes a {{touch(BlockCacheKey)}} operation.
> * L1 hits in {{TieredInclusiveTopology}} are propagated to L2.
> * L2 can update its access metadata without fetching the cached block.
> * {{TieredExclusiveTopology}} and {{SingleTierTopology}} do not generate
> unnecessary cross-tier touches.
> * {{touch()}} does not affect cache hit/miss statistics.
> * Existing {{CacheEngine}} implementations remain compatible through the
> default no-op implementation.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
