HBASE-29669 Implement basic row cache by EungsopYoo · Pull Request #7901 · apache/hbase

EungsopYoo · 2026-03-11T07:58:29Z

No description provided.

EungsopYoo · 2026-03-11T08:05:15Z

@wchevreuil
I have completed the implementation of the basic row cache. Please review it.

By the way, it seems like CI is not working, do you know why?

liuxiaocs7 · 2026-03-16T03:45:23Z

CI has been migrated from Jenkins to GitHub Actions, and the pre-commit GitHub Jenkins job should have been disabled, more details: https://lists.apache.org/thread/z6o5hhsd9goh5j7fcl4bnwzsktlwlwl4

Some previously created feature branches also have this issue; perhaps we could give it a try, such as: https://lists.apache.org/thread/hrl7dktkrb8zwwbv62my68lvtbplv7sr

Thanks!

liuxiaocs7 · 2026-03-16T03:47:09Z

+
+  void removeTableLevelBarrier(HRegion region) {
+    regionLevelBarrierMap.computeIfPresent(region, (k, counter) -> {
+      int remaining = counter.decrementAndGet();
+      return (remaining <= 0) ? null : counter;
+    });
+  }


this method should be renamed by removeRegionLevelBarrier?

Done, renamed to removeRegionLevelBarrier. (95de81b)

liuxiaocs7 · 2026-03-16T03:48:17Z

+  // For testing only
+  void setRowCache(RowCache rowCache) {
+    this.rowCache = rowCache;
+  }


For testing only, we could use @RestrictedApi annotation?

Done, added @RestrictedApi annotation. (267f4ef)

liuxiaocs7 · 2026-03-16T03:49:14Z

  private BlockCache l2Cache = null;
  private MobFileCache mobFileCache;
  private CacheStats cacheStats;
+  private final RowCache rowCache;


better move to L71 under MobFileCache?

Done, moved rowCache field next to MobFileCache. (782dd75)

liuxiaocs7 · 2026-03-16T03:50:28Z

+    RSRpcServices rsRpcServices = this.regionServer.getRSRpcServices();
+    this.rowCache = rsRpcServices == null ? null : rsRpcServices.getServer().getRowCache();


better to use a method initRowCache() here, just like blockcache and mobfilecache does?

Done, extracted initRowCache() method. (782dd75)

liuxiaocs7 · 2026-03-16T03:51:14Z

+      // The row cache for the region has been enabled again
+      rowCache.removeTableLevelBarrier(region);


ditto region level

Done, renamed to removeRegionLevelBarrier. (95de81b)

liuxiaocs7 · 2026-03-16T03:52:24Z

+@Category({ RegionServerTests.class, MediumTests.class })
+public class TestRowCache {
+  @ClassRule
+  public static final HBaseClassTestRule CLASS_RULE =
+    HBaseClassTestRule.forClass(TestRowCache.class);


For new tests, it is best to use JUnit5 directly, since existing unit tests are currently being migrated to JUnit5.

Done, migrated TestRowCache to JUnit5. (1fc9d5c)

EungsopYoo · 2026-04-04T06:54:32Z

@liuxiaocs7
Sorry, I just noticed your comments. I’ll take a look and follow up early next week.

…ethod

wchevreuil · 2026-04-27T14:44:38Z

-    // TODO: implement row cache
-    rowCacheStrategy = null;
+    // Currently we only support TinyLfu implementation
+    rowCacheStrategy = new TinyLfuRowCacheStrategy(MemorySizeUtil.getRowCacheSize(conf));


Should be made pluggable.

Done. Introduced row.cache.strategy.class config key so a custom RowCacheStrategy can be plugged in via configuration; default remains TinyLfuRowCacheStrategy. This follows the same convention used by MemStore and RegionSplitPolicy (Configuration-arg constructor + ReflectionUtils.newInstance).

Commit: 6739f06

wchevreuil · 2026-04-27T16:19:49Z

-    } catch (CloneNotSupportedException ignored) {
-      // Not able to cache row cells, ignore
-    }
+  void populateCache(HRegion region, List<Cell> results, RowCacheKey key) {


nit: just call this method "cache".

Do we really need the regionLevelBarrierMap key type to be HRegion, or could it be just the String encoded region name? That way, we don't need an extra HRegion parameter in this method.

Done.

Renamed populateCache to cache.

Changed regionLevelBarrierMap key type from HRegion to the encoded region name (String). The encoded name is the canonical region identifier already used elsewhere (e.g., RowCacheKey.isSameRegion). As a result, cache no longer needs an HRegion parameter; it derives the encoded region name from the RowCacheKey.

I kept the external signatures of create/remove/getRegionLevelBarrier taking HRegion to keep the caller's intent explicit; only the internal map key type changes. Let me know if you'd prefer those to take String as well.

Commit: e985135

wchevreuil · 2026-04-27T17:04:41Z

+      // Put with TTL is not allowed on tables with row cache enabled, because cached rows cannot
+      // track TTL expiration
+      if (isRowCacheEnabled) {
+        if (put.getTTL() != Long.MAX_VALUE) {
+          throw new DoNotRetryIOException(
+            "Tables with row cache enabled do not allow setting TTL on Puts");
+        }
+      }


Can't we simply implement TTL expiration within the RowCache.getRow logic?

Good point — turned out to be feasible. Done.

Quick verification: I was initially worried that Get results don't carry TTL tags (the design doc assumed this), but that's actually a client-side property — the RPC codec strips tags. Server-side cells delivered to RowCache preserve their TTL tag (carried forward by TagUtil.carryForwardTTLTag during mutation, the same tag ScanQueryMatcher.isCellTTLExpired reads). Confirmed with a probe test against a live region.

Implementation:

RowCells now precomputes the earliest TTL expiration time across its cells during construction. Cells without a TTL tag yield Long.MAX_VALUE, so RowCells.isExpired(now) is an O(1) long comparison on every cache hit — no per-cell tag iteration on the hot path.

RowCache.tryGetFromCache: if the row is expired, evict and fall back to the storage read path.

RowCache.cache: skip caching when results are empty (otherwise an evicted-then-empty row would be re-cached as an empty entry).

HRegion.put: removed the guard that rejected Puts with TTL.

CF-level TTL still disables the row cache via canCacheRow's isDefaultTtl check; that policy is unchanged for now.

Commit: d8fef38

Introduce row.cache.strategy.class configuration key to allow operators to plug in custom RowCacheStrategy implementations. The default remains TinyLfuRowCacheStrategy. RowCacheStrategy implementations must now provide a public constructor that takes a Configuration argument, following the same convention used by MemStore and RegionSplitPolicy.

Address review feedback: - Rename RowCache.populateCache to cache. - Change regionLevelBarrierMap key type from HRegion to the encoded region name (String). The encoded name is the canonical region identifier already used elsewhere (e.g., RowCacheKey.isSameRegion). - The cache method no longer needs an HRegion parameter; it derives the encoded region name from the RowCacheKey. The external signatures of create/remove/getRegionLevelBarrier still take HRegion to make the caller's intent explicit; only the internal map key type changes.

…n hit Address review feedback: instead of rejecting Puts with TTL on row cache-enabled tables, check TTL expiration when serving cache hits. - RowCells: precompute the earliest TTL expiration time across the contained cells during construction, exposing isExpired(now) for an O(1) check on each cache hit. Cells without a TTL tag yield Long.MAX_VALUE so the check short-circuits. - RowCache.tryGetFromCache: if the cached row is expired, evict it and fall back to the storage read path. - RowCache.cache: skip caching when results are empty. - HRegion.put: remove the guard that rejected Puts with TTL on row cache-enabled tables. Server-side cells preserve their TTL tag (carried forward by TagUtil.carryForwardTTLTag during mutation), so the same expiration check used by ScanQueryMatcher can be applied at the cache layer. CF-level TTL still disables the row cache via canCacheRow's isDefaultTtl check; that policy is unchanged.

The CI workflows invoke scripts under dev-support/ that were missing or outdated on this branch, causing GHA runs to fail at startup.

Implement basic row cache

61c0798

liuxiaocs7 reviewed Mar 16, 2026

View reviewed changes

Add GitHub Actions CI workflow files from master

cd93c2f

EungsopYoo force-pushed the HBASE-29669 branch from 819338e to cd93c2f Compare April 6, 2026 00:16

terence-yoo added 4 commits April 6, 2026 09:53

Rename removeTableLevelBarrier to removeRegionLevelBarrier

95de81b

Move rowCache field next to MobFileCache and extract initRowCache() m…

782dd75

…ethod

Add @RestrictedApi annotation to setRowCache() in HRegion

267f4ef

Migrate TestRowCache from JUnit4 to JUnit5

1fc9d5c

wchevreuil reviewed Apr 27, 2026

View reviewed changes

terence-yoo added 4 commits April 30, 2026 20:59

Sync dev-support scripts from master

8b89b7c

The CI workflows invoke scripts under dev-support/ that were missing or outdated on this branch, causing GHA runs to fail at startup.

		RSRpcServices rsRpcServices = this.regionServer.getRSRpcServices();
		this.rowCache = rsRpcServices == null ? null : rsRpcServices.getServer().getRowCache();

		// The row cache for the region has been enabled again
		rowCache.removeTableLevelBarrier(region);

Conversation

EungsopYoo commented Mar 11, 2026

Uh oh!

EungsopYoo commented Mar 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

liuxiaocs7 commented Mar 16, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

EungsopYoo Apr 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

EungsopYoo Apr 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

EungsopYoo Apr 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

EungsopYoo Apr 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

EungsopYoo Apr 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

liuxiaocs7 Mar 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

EungsopYoo Apr 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

EungsopYoo commented Apr 4, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

EungsopYoo commented Mar 11, 2026 •

edited

Loading

EungsopYoo Apr 6, 2026 •

edited

Loading

EungsopYoo Apr 6, 2026 •

edited

Loading

EungsopYoo Apr 6, 2026 •

edited

Loading

EungsopYoo Apr 6, 2026 •

edited

Loading

EungsopYoo Apr 6, 2026 •

edited

Loading

liuxiaocs7 Mar 16, 2026 •

edited

Loading

EungsopYoo Apr 6, 2026 •

edited

Loading