Repository navigation
[improve][broker] PIP-379: Remove the classic Shared and Key_Shared dispatcher implementations - #26687
Merged
Conversation
… in tests ### Motivation The classic Shared/Key_Shared dispatchers are being removed in 5.0. Before the implementation can go, the tests that parameterize on it have to stop doing so. `KeySharedImplementationType` ran the four Key_Shared test classes twice via a `@Factory`, starting two brokers for the largest of them. ### Modifications - Delete `org.apache.pulsar.tests.KeySharedImplementationType` and unwind its four users: the `@Factory`, the extra constructors, the inject-only `currentImplementationType` data provider, the `prependImplementationTypeToData` wrapping of the real providers, and the classic config setters. - `KeySharedSubscriptionTest.testContinueDispatchMessagesWhenMessageTTL`: the `!impl.classic` guards collapse to the modern form. Note the consumer3 check at the second site becomes unconditional, matching the already-unconditional consumer2 check above it. - `testReadAheadLimit` and `testOrderingAfterReconnects` lose their `skipIfClassic()` and now always run. - `KeySharedSubscriptionMaxUnackedMessagesTest` keeps its `KeySharedSelectorType` enum including `AutoSplit_Classic` — that constant selects the hash-range *selector* (`subscriptionKeySharedUseConsistentHashing`), an orthogonal axis. No production code changes.
…assic implementation ### Motivation Follow-up to the Key_Shared test cleanup, and the last step before the classic Shared/Key_Shared dispatchers can be deleted. Four broker test classes still ran every case twice, once against each dispatcher implementation. ### Modifications - `ConsumerStatsTest`: `classicAndSubscriptionType` becomes `sharedSubscriptionTypes` (`Shared`, `Key_Shared`); `testConsumerStatsOutput` drops the `classicDispatchers` parameter and the config setters, and its expected-field set keeps only the modern `drainingHashes` / `keyHashRangeArrays` arm. - `SharedDispatcherPermitAccountingTest`: `dispatcherImplementations` is gone and `flowRaceDispatcherVariants` collapses to the two subscription types; `createTestContext` loses its `classic` argument (the unused 1-arg overload is dropped) and the dispatcher branches, instance-of assertions and the two `totalAvailablePermits` helpers unwrap to the modern dispatcher. `untrackedSubscriptionVariants` is untouched — its first column is `persistent`, not `classic`. - `SharedSubscriptionUnackedMessagesAccountingTest`: `@Factory`, the `classic` field and constructor go; one broker instead of two. - `PersistentDispatcherMultipleConsumersTest`: `readConflationDispatcherTypes` is removed and `createReadConflationDispatcher` always builds the modern dispatcher. No production code changes.
… implementations ### Motivation PIP-379 replaced the Key_Shared "recently joined consumers" machinery with the draining-hashes tracker. The implementation shipped in 4.0 LTS but kept the pre-4.0 dispatchers alive behind two feature flags as a rollback escape hatch: - `subscriptionKeySharedUseClassicPersistentImplementation` - `subscriptionSharedUseClassicPersistentImplementation` Both default to `false` and neither appears in `conf/broker.conf` or `conf/standalone.conf`. Two LTS lines later the escape hatch has outlived its purpose: ~2000 lines of duplicated dispatcher logic, a classic/modern branch threaded through `PersistentSubscription`, and a path that lacks read-ahead limiting while carrying the ordering defects (apache#23307, apache#21199) PIP-379 exists to fix. 5.0 is the right major release to delete it. ### Modifications - Delete `PersistentDispatcherMultipleConsumersClassic` and `PersistentStickyKeyDispatcherMultipleConsumersClassic`, their two dedicated test classes, and `NonEntryCacheKeySharedSubscriptionV30Test` (it forces the classic implementation and drives the classic-only `setSortRecentlyJoinedConsumersIfNeeded` / `getRecentlyJoinedConsumers` hooks). - Delete both `ServiceConfiguration` flags. - `PersistentSubscription`: the Shared branch collapses to `PersistentDispatcherMultipleConsumers`, and the Key_Shared branch loses only the classic arm — the PIP-486 `ksm.isEntryBucketDispatch()` arm stays. - Drop the two commented-out `PULSAR_PREFIX_...` lines in `AbstractPulsarProfilingTest`. A stale `subscriptionKeySharedUseClassicPersistentImplementation=true` left in `broker.conf` is silently ignored — `FieldParser.update` iterates declared fields, never the properties — and surfaces only in the `configOverrides` attribute of the startup log. Both flags were `dynamic = true`, so a stale value may sit in `/admin/configuration`; `BrokerService` already logs and skips unknown dynamic keys. The PIP-322 classic *dispatch rate limiter* is a different feature that first shipped in 4.1.0 and is deliberately out of scope here.
…lementations used ### Motivation With the classic dispatchers gone, the seams that existed solely to distinguish them from the modern ones are dead weight. ### Modifications - `isClassic()` is removed at all four sites: the abstract declaration on `AbstractPersistentDispatcherMultipleConsumers`, the `StickyKeyDispatcher` interface method, the `return false` on `PersistentDispatcherMultipleConsumers`, and the stats branch in `PersistentSubscription`, which now always emits `keyHashRangeArrays`. It was never called through the base type anywhere, including tests. - `StickyKeyDispatcher.getRecentlyJoinedConsumers()` — a `default` returning `null` whose only overrider was the classic dispatcher — is removed together with the block in `PersistentSubscription` that fed `subStats.consumersAfterMarkDeletePosition` from it. - `MessageRedeliveryController`: the `isClassicDispatcher` field and the 2-arg constructor are gone, `add()` loses its classic branches, and the write-only `containsStickyKeyHashes(Set<Integer>)` overload is removed in favour of the singular `containsStickyKeyHash(int)`. - `MessageRedeliveryControllerTest`: drop `testClassicOutOfOrderDoesNotInitializePositionHashMap` and rewrite `testContainsStickyKeyHashes` onto the singular form. The plural assertions were an any-match loop, so `containsStickyKeyHashes(Set.of(104, 105))` passed on 104 alone; the singular version now asserts 104 present and 105 absent. - Reword the stale "recent joined consumer tracking" comment in `PersistentStickyKeyDispatcherMultipleConsumers` — it describes what is actually the `drainingHashesRequired` flag. `AbstractPersistentDispatcherMultipleConsumers` is deliberately kept: it is the parameter type of the user-pluggable `DelayedDeliveryTrackerFactory.newTracker(...)` SPI, the narrowing type `BrokerService` uses for `blockedDispatchers`, and the type seven test files mock.
…ats fields ### Motivation With the classic dispatchers gone, three stats fields have no producer left. This is the only commit in the series with a public admin-API break, so it is kept as one reviewable diff. - `ConsumerStats.getReadPositionWhenJoining()` and `getKeyHashRanges()` have been `@Deprecated` since apache#23429 (Oct 2024). They shipped deprecated in 4.0.0 and had 4.0, 4.1 and 4.2 to warn. - `SubscriptionStats.getConsumersAfterMarkDeletePosition()` was never marked `@Deprecated`, although `pip/pip-379.md` already declared it removed. ### Modifications Remove all three from the admin API interfaces, the `*Impl` classes (field, `add()`, `clear()`, getter) and `Consumer` (the `readPositionWhenJoining` field, its `updateStats` branch, and `setReadPositionWhenJoining`, whose only caller was the classic dispatcher). ### JSON-visible change `consumersAfterMarkDeletePosition` was initialized to an empty `LinkedHashMap`, so persistent subscriptions emit `"consumersAfterMarkDeletePosition": {}` today and will stop. The other two are `null` on the modern path and the admin mapper uses `Include.NON_NULL`, so their removal is JSON-invisible. The disappearing `{}` is a deliberate 5.0 stats-shape change. ### Why `consumersAfterMarkDeletePosition` is removed, not renamed It was never a general "consumers past the mark-delete position" statistic that some other mechanism could repopulate — it was the classic dispatcher's recently-joined-consumers map under a name that described its eviction rule: - **Its own javadoc contradicts its name.** Both `SubscriptionStats` and `SubscriptionStatsImpl` carried the line *"This is for Key_Shared subscription to get the recentJoinedConsumers in the Key_Shared subscription."* The documented subject is `recentJoinedConsumers`. - **The name describes a retention rule, not the data.** The key is a consumer name; the value is that consumer's cursor read position at the moment it joined — the same value surfaced as the consumer-level `readPositionWhenJoining`. "After mark delete position" referred only to the pruning invariant in `removeConsumersFromRecentJoinedConsumers()`, which dropped an entry once the mark-delete position passed the join position. The name conveys neither the key nor the value, only an incidental eviction condition. - **It had no meaning outside the feature.** Its sole writer was `PersistentSubscription`, fed exclusively by `StickyKeyDispatcher.getRecentlyJoinedConsumers()`, whose only overrider was the classic dispatcher. PIP-282 renamed the consumer-level sibling (`readPositionWhenJoining` → `lastSentPositionWhenJoining`) but left this one alone, and PIP-379 then declared it removed outright. `ConsumerStatsImpl` keeps its class-level `@SuppressWarnings("deprecation")` — `lastAckedTimestamp` and `lastConsumedTimestamp` remain deprecated.
lhotari
requested review from
Technoboy-,
dao-jun,
david-streamlio,
merlimat and
nodece
September 22, 2026 13:20
void-ptr974
approved these changes
Sep 22, 2026
dao-jun
approved these changes
Sep 22, 2026
merlimat
approved these changes
Sep 22, 2026
lhotari
pushed a commit
that referenced
this pull request
Oct 2, 2026
…ch index acks (#26809) (cherry picked from commit 1046481) [branch-4.x] Also applied to the classic Shared and Key_Shared dispatchers (PersistentDispatcherMultipleConsumersClassic, PersistentStickyKeyDispatcherMultipleConsumersClassic), which were removed on master by #26687. The test covers both implementations.
lhotari
pushed a commit
that referenced
this pull request
Oct 2, 2026
…ch index acks (#26809) (cherry picked from commit 1046481) [branch-4.x] Also applied to the classic Shared and Key_Shared dispatchers (PersistentDispatcherMultipleConsumersClassic, PersistentStickyKeyDispatcherMultipleConsumersClassic), which were removed on master by #26687. The test covers both implementations.
Radiancebobo
pushed a commit
to Radiancebobo/pulsar
that referenced
this pull request
Oct 8, 2026
…ispatcher implementations (apache#26687)
1 of 11 tasks
crossoverJie
added a commit
to crossoverJie/pulsar-client-go
that referenced
this pull request
Oct 15, 2026
apache/pulsar#26687 (PIP-379: Remove the classic Shared and Key_Shared dispatcher implementations, shipped in 5.0.0) removed consumersAfterMarkDeletePosition from the subscription stats. The broker no longer serializes the field, so it unmarshals to a nil map instead of the empty map the test asserted. Assert the map is empty, which holds both for the old behavior (empty map) and the new one (absent/nil). Verified against Pulsar 4.2.4, 4.2.5 and 5.0.0. Co-Authored-By: CodeBuddy Code <noreply@codebuddy.ai>
5 of 14 tasks
Asthenia0412
pushed a commit
to Asthenia0412/pulsar
that referenced
this pull request
Oct 17, 2026
…sage Issue: apache#26395 Motivation: A Shared subscription stops dispatching forever when it runs into a chunk of a chunked message whose chunk 0 is no longer in the uuidToConsumer cache. This happens after topic unload, broker restart, or consumer disconnect. SharedConsumerAssignor pins all chunks of a message to one consumer through the in-memory uuidToConsumer map, which is only written when chunk 0 is assigned. The original code returned null for any non-zero chunk with no cached mapping: if (metadata.getChunkId() != 0) { // Not the first chunk, skip it return null; } The null return causes the entry to be handed to the replay queue. The dispatcher drains the replay queue before every normal read, so it keeps replaying the same orphan chunk forever and the mark-delete position never moves. The same stall occurs when a consumer is removed while chunks are still pinned to it — the mapping survives, the disconnected consumer has no permits, so every remaining chunk also gets replayed forever. Modifications: - SharedConsumerAssignor.getConsumerForUuid(): orphan chunks (chunkId != 0 with no cached mapping) are now dispatched to the default consumer instead of returning null. The chunk 0 is what seeds the mapping, and the cursor always re-reads chunk 0 ahead of later chunks while it is still unacked, so a missing mapping means chunk 0 is already acked and the message can never be reassembled. Dispatching it lets the client discard the incomplete chunked message and move on. This path is logged at WARN so operators can detect silently-acked chunks. - SharedConsumerAssignor.removeConsumer(Consumer): new public method that drops all uuid mappings pointing to a removed consumer. Called from the dispatcher when a consumer disconnects. - SharedConsumerAssignor.clear(): new public method to drop all uuid mappings. Called when all consumers are removed from the dispatcher to prevent leaks from messages whose last chunk is never published. - PersistentDispatcherMultipleConsumers: call assignor.removeConsumer() in removeConsumer() and assignor.clear() in clearComponentsAfterRemovedAllConsumers(). PersistentStickyKeyDispatcherMultipleConsumers inherits both via super(), so the fix covers Key_Shared subscriptions too. Note on PR apache#26396: this supersedes PR apache#26396, which has a merge conflict because it modifies PersistentDispatcherMultipleConsumersClassic.java — a file that was removed by PIP-379 (apache#26687) on 2026-09-22. The core fix in apache#26396 is correct; this PR applies the same logic to the file set that exists on master today. Testing: - SharedConsumerAssignorTest.testOrphanChunkIsDispatchedInsteadOfReplayedForever: verifies orphan chunks are dispatched, not replayed, and do not create a uuid mapping. Fails on original code. - SharedConsumerAssignorTest.testUuidMappingIsRemovedWhenConsumerIsRemoved: verifies the uuid mapping is cleaned up when its consumer is removed, and the remaining chunk can be dispatched to another consumer. - SharedConsumerAssignorTest.testClearRemovesAllUuidMappings: verifies the map is cleared properly to prevent leaks. - All existing SharedConsumerAssignorTest tests pass. - All existing PersistentDispatcherMultipleConsumersTest tests pass.
1 of 10 tasks
merlimat
added a commit
to void-ptr974/pulsar
that referenced
this pull request
Oct 18, 2026
- Drop the PersistentDispatcherMultipleConsumersClassic changes and tests: the classic dispatcher was removed in apache#26687. - Re-port the async snapshot recovery onto the BucketContext / index code from apache#26251, keeping the recovery synchronized on the tracker. - Keep the dispatcher's delayed-delivery stats reads lock-free (apache#25990). - A failed tracker creation keeps the delayed messages pending and dispatch paused, and retries with a backoff, instead of installing DelayedDeliveryTracker.DISABLE for the life of the dispatcher. - A closed dispatcher doesn't start a tracker creation. - Handing the pending messages to the new tracker is exception safe: the creation still completes, the remaining messages are replayed and dispatch resumes. - The bucket factory falls back to the in-memory tracker on any async recovery failure. - Tracker creation logs at DEBUG; only failures and a creation still pending after 30 s log a WARN. - Tests verify closeAsync(), cover the closed dispatcher, elapsed messages, a throwing tracker and the retry, check that the dispatcher lock is free during recovery, and restore the shared broker's topic map.
Asthenia0412
pushed a commit
to Asthenia0412/pulsar
that referenced
this pull request
Oct 18, 2026
…sage Issue: apache#26395 Motivation: A Shared subscription stops dispatching forever when it runs into a chunk of a chunked message whose chunk 0 is no longer in the uuidToConsumer cache. This happens after topic unload, broker restart, or consumer disconnect. SharedConsumerAssignor pins all chunks of a message to one consumer through the in-memory uuidToConsumer map, which is only written when chunk 0 is assigned. The original code returned null for any non-zero chunk with no cached mapping: if (metadata.getChunkId() != 0) { // Not the first chunk, skip it return null; } The null return causes the entry to be handed to the replay queue. The dispatcher drains the replay queue before every normal read, so it keeps replaying the same orphan chunk forever and the mark-delete position never moves. The same stall occurs when a consumer is removed while chunks are still pinned to it — the mapping survives, the disconnected consumer has no permits, so every remaining chunk also gets replayed forever. Modifications: - SharedConsumerAssignor.getConsumerForUuid(): orphan chunks (chunkId != 0 with no cached mapping) are now dispatched to the default consumer instead of returning null. The chunk 0 is what seeds the mapping, and the cursor always re-reads chunk 0 ahead of later chunks while it is still unacked, so a missing mapping means chunk 0 is already acked and the message can never be reassembled. Dispatching it lets the client discard the incomplete chunked message and move on. This path is logged at WARN so operators can detect silently-acked chunks. - SharedConsumerAssignor.removeConsumer(Consumer): new public method that drops all uuid mappings pointing to a removed consumer. Called from the dispatcher when a consumer disconnects. - SharedConsumerAssignor.clear(): new public method to drop all uuid mappings. Called when all consumers are removed from the dispatcher to prevent leaks from messages whose last chunk is never published. - PersistentDispatcherMultipleConsumers: call assignor.removeConsumer() in removeConsumer() and assignor.clear() in clearComponentsAfterRemovedAllConsumers(). PersistentStickyKeyDispatcherMultipleConsumers inherits both via super(), so the fix covers Key_Shared subscriptions too. Note on PR apache#26396: this supersedes PR apache#26396, which has a merge conflict because it modifies PersistentDispatcherMultipleConsumersClassic.java — a file that was removed by PIP-379 (apache#26687) on 2026-09-22. The core fix in apache#26396 is correct; this PR applies the same logic to the file set that exists on master today. Testing: - SharedConsumerAssignorTest.testOrphanChunkIsDispatchedInsteadOfReplayedForever: verifies orphan chunks are dispatched, not replayed, and do not create a uuid mapping. Fails on original code. - SharedConsumerAssignorTest.testUuidMappingIsRemovedWhenConsumerIsRemoved: verifies the uuid mapping is cleaned up when its consumer is removed, and the remaining chunk can be dispatched to another consumer. - SharedConsumerAssignorTest.testClearRemovesAllUuidMappings: verifies the map is cleared properly to prevent leaks. - All existing SharedConsumerAssignorTest tests pass. - All existing PersistentDispatcherMultipleConsumersTest tests pass.
nodece
pushed a commit
to apache/pulsar-client-go
that referenced
this pull request
Oct 20, 2026
…ior changes (#1545) * [fix][test] Make TestSetTopicAutoCreation tolerant of broker error message changes The test asserted exact broker error strings, which broke when Pulsar changed the messages: - apache/pulsar#25163 (shipped in 4.2.5) relaxed the non-partitioned autoTopicCreation override validation and changed the rejection text from "[defaultNumPartitions] is not allowed to be set when the type is non-partition." to "[defaultNumPartitions] must be null, 0 or 1 when the type is non-partitioned." - 5.0.0 changed the missing-tenant error from "Tenant does not exist" to "Namespace does not exist". Replace the single exact-match `errReason` with an `errReasons` slice and assert containment, so the test passes against both old and new broker versions instead of being brittle to broker wording changes. Verified against Pulsar 4.2.4, 4.2.5 and 5.0.0. Co-Authored-By: CodeBuddy Code <noreply@codebuddy.ai> * [fix][test] Authenticate transaction metadata request as super-user TestTxn_TCClient read transaction metadata through the raw admin HTTP helper (httpGet -> httpDo), which calls the plain-text web service without any credentials, so the broker authenticates it as the "anonymous" role. apache/pulsar#26773 ("Improve transaction admin checks", shipped in 4.2.5 and 5.0.0) added validateSuperUserAccess() to the transaction metadata admin endpoint (GET /admin/v3/transactions/transactionMetadata/{mostSigBits}/{leastSigBits}). The unauthenticated request is now rejected with HTTP 401, so transactionStats fails and the test panics on the nil stats map. Add an httpGetWithTLS helper that calls the TLS web service using the test client certificate, whose CN ("admin") is a super-user role in standalone.conf, and use it for the transaction metadata request. Verified against Pulsar 4.2.4, 4.2.5 and 5.0.0. Co-Authored-By: CodeBuddy Code <noreply@codebuddy.ai> * [fix][test] Make TestPartitionState tolerant of removed stats field apache/pulsar#26687 (PIP-379: Remove the classic Shared and Key_Shared dispatcher implementations, shipped in 5.0.0) removed consumersAfterMarkDeletePosition from the subscription stats. The broker no longer serializes the field, so it unmarshals to a nil map instead of the empty map the test asserted. Assert the map is empty, which holds both for the old behavior (empty map) and the new one (absent/nil). Verified against Pulsar 4.2.4, 4.2.5 and 5.0.0. Co-Authored-By: CodeBuddy Code <noreply@codebuddy.ai> * [fix][test] Address review feedback on admin test helpers - httpGetWithTLS: give the HTTPS test client a 30s timeout so a stalled broker/network path cannot hang the test run. - TestSetTopicAutoCreation: replace the unchecked `err.(rest.Error)` type assertion (which panics on transport/TLS errors) with `require.ErrorAs`. Co-Authored-By: CodeBuddy Code <noreply@codebuddy.ai> * [fix][test] Cache the TLS admin HTTP client in test helper Build the HTTPS client (key pair, CA pool) once via sync.Once and reuse it across calls instead of rebuilding it on every request. Co-Authored-By: CodeBuddy Code <noreply@codebuddy.ai> --------- Co-authored-by: CodeBuddy Code <noreply@codebuddy.ai>
This was referenced Oct 21, 2026
Closed
nodece
pushed a commit
to ascentstream/pulsar
that referenced
this pull request
Nov 6, 2026
…ch index acks (apache#26809) (cherry picked from commit 1046481) [branch-4.x] Also applied to the classic Shared and Key_Shared dispatchers (PersistentDispatcherMultipleConsumersClassic, PersistentStickyKeyDispatcherMultipleConsumersClassic), which were removed on master by apache#26687. The test covers both implementations.
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
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.
PIP: 379 (
pip/pip-379.md)Motivation
PIP-379 replaced the Key_Shared "recently joined consumers" machinery with the draining-hashes
tracker, and the matching work replaced the Shared dispatcher. Both landed in 4.0.0 LTS
(Oct 2024) and became the default in that release. The pre-4.0 implementations were kept alive
behind two feature flags purely as a rollback escape hatch:
subscriptionKeySharedUseClassicPersistentImplementationsubscriptionSharedUseClassicPersistentImplementation5.0 is the right release to delete them.
The classic path was already an escape hatch, not a supported alternative
It shipped in 4.0.0 with every signal of a temporary fallback:
false, so no default deployment has used the classic path since 4.0.0.conf/broker.conforconf/standalone.conf— they were never advertisedas a supported knob.
ConsumerStats.getReadPositionWhenJoining()andgetKeyHashRanges()were marked@Deprecatedin [improve][broker][PIP-379] Add observability stats for "draining hashes"Â #23429 back in Oct 2024, i.e. they shippeddeprecated in 4.0.0 and have carried that warning through 4.0, 4.1 and 4.2.
pip/pip-379.mdalready listsconsumersAfterMarkDeletePositionas removed.Users have therefore had two full LTS lines of notice, with the modern implementation as the
default the whole time.
PIP-379 has been the default for two years; the classic path still has the original defects
The draining-hashes implementation has been the default for every deployment on 4.0+ for roughly
two years, across three release lines. The classic implementation still has the problems PIP-379
documents in its Motivation section:
during consumer changes (#23307);
is blocked, even when other keys could be dispatched independently;
drainingHashes,drainingHashesCount,drainingHashesClearedTotalanddrainingHashesUnackedMessages;It also carries #21199 (Key_Shared subscription
gets stuck after consumer reconnects). Keeping the flag keeps a supported way to opt back into
these defects.
The classic path is excluded from the PIP-430 cache improvements
PIP-430's expected-read-count eviction (
cacheEvictionByExpectedReadCount, defaulttrue)relies on the dispatcher incrementing an entry's expected read count when that entry is pushed onto
the replay queue, so the entry is de-prioritised for size-based eviction and is more likely to still
be cached when the Key_Shared consumer can finally take it (PIP-430, "Detailed Design").
That call —
entry.getReadCountHandler().incrementExpectedReadCount()— exists only inPersistentStickyKeyDispatcherMultipleConsumers. The classic dispatcher never had it. So on 5.0 auser who sets the classic flag would silently get materially worse cache hit rates on exactly the
replay traffic the cache improvement was designed for, with nothing pointing at the flag as the
cause. Features built on top of PIP-379 will keep widening this gap.
Maintenance cost, concretely
The two classes are ~2000 lines of duplicated dispatcher logic, plus a classic/modern branch
threaded through
PersistentSubscription,MessageRedeliveryControllerand the admin statssurface, and a
@Factorythat started a second broker for the largest Key_Shared test class.The clearest evidence is the commit this branch is based on:
#26677 "Bound classic Key_Shared dispatcher replay
queue look-ahead". It retrofits a look-ahead bound onto the classic dispatcher by reaching into the
modern class for
PersistentStickyKeyDispatcherMultipleConsumers.getEffectiveLookAheadLimit(...),and even then only approximates it — its own comment concedes "a bounded look-ahead limit with an
acceptable one-batch overflow". That is the ongoing tax: a capability designed for the modern
dispatcher has to be re-derived, approximately, for a path nobody runs by default, and every future
fix in this area faces the same choice between doing the work twice or leaving the two paths
divergent.
Deleting it removes that tax for the whole 5.0 line.
Modifications
Five commits, ordered so the tree compiles and tests pass at each one. Deletion cannot come first:
eight test files call the config setters, six name the classic classes, and
ConsumerStatsTestasserts an exact JSON field set — a runtime dependency javac would not catch. The tests are
therefore de-parameterized first.
[improve][test]Stop exercising the classic Key_Shared implementation in tests — deleteorg.apache.pulsar.tests.KeySharedImplementationTypeand unwind its four users (the@Factory,the extra constructors, the inject-only
currentImplementationTypeprovider, and theprependImplementationTypeToDatawrapping).testReadAheadLimitandtestOrderingAfterReconnectslose theirskipIfClassic()and now always run.[improve][test]Stop parameterizing Shared-dispatcher tests on the classic implementation —ConsumerStatsTest,SharedDispatcherPermitAccountingTest,SharedSubscriptionUnackedMessagesAccountingTestandPersistentDispatcherMultipleConsumersTest.[improve][broker]Remove the classic Shared and Key_Shared dispatcher implementations —delete
PersistentDispatcherMultipleConsumersClassic,PersistentStickyKeyDispatcherMultipleConsumersClassic, their dedicated tests andNonEntryCacheKeySharedSubscriptionV30Test; delete bothServiceConfigurationflags; collapsethe branches in
PersistentSubscription.[improve][broker]Drop the dispatcher seams that only the classic implementations used —isClassic(),StickyKeyDispatcher.getRecentlyJoinedConsumers(), and theMessageRedeliveryControllerisClassicDispatcherfield, 2-arg constructor andcontainsStickyKeyHashes(Set<Integer>)overload.[improve][admin]Remove the classic-only consumer and subscription stats fields — the onlycommit with a public API break, kept as one reviewable diff. See Upgrade impact.
Net: 28 files, +208 / −4031 (production code: +13 / −2173).
The PIP-486
ksm.isEntryBucketDispatch()arm inPersistentSubscriptionis kept, as isAbstractPersistentDispatcherMultipleConsumers— it is the parameter type of the user-pluggableDelayedDeliveryTrackerFactory.newTracker(...)SPI, the narrowing typeBrokerServiceuses forblockedDispatchers, and the type seven test files mock.Deliberately out of scope: the PIP-322 classic dispatch rate limiter
(
DispatchRateLimiterClassicImpl,DispatchRateLimiterFactoryClassic,dispatchRateLimiterFactoryClassName). Different feature, first shipped in 4.1.0 — a much shorterdeprecation life. Separate PR.
Upgrade impact
Stale config is ignored, not rejected. A leftover
subscriptionKeySharedUseClassicPersistentImplementation=trueinbroker.confdoes not failstartup —
FieldParser.updateiterates declared fields, never the properties — and surfaces only inthe
configOverridesattribute of the "Messaging service is ready" startup log. Both flags weredynamic = true, so a stale value may also sit in/admin/configuration;BrokerServicealreadylogs and skips unknown dynamic keys. The broker simply uses the modern dispatcher, which is what it
already did for every default deployment since 4.0.0.
One JSON-visible stats change.
SubscriptionStats.consumersAfterMarkDeletePositionwasinitialized to an empty
LinkedHashMap, so persistent subscriptions emit"consumersAfterMarkDeletePosition": {}today and will stop emitting it. The other two removedfields are
nullon the modern path and the admin mapper usesInclude.NON_NULL, so their removalis JSON-invisible. None of the three is exposed via Prometheus metrics. This is a deliberate 5.0
stats-shape change.
Why
consumersAfterMarkDeletePositionis removed rather than renamed or kept as an empty map.It was never a general "consumers past the mark-delete position" statistic that some other mechanism
could repopulate — it was the classic dispatcher's recently-joined-consumers map under a name that
described its eviction rule:
SubscriptionStatsandSubscriptionStatsImplcarried the line "This is for Key_Shared subscription to get the recentJoinedConsumers in the
Key_Shared subscription." The documented subject is
recentJoinedConsumers.consumer's cursor read position at the moment it joined — the same value surfaced as the
consumer-level
readPositionWhenJoining. "After mark delete position" referred only to thepruning invariant in
removeConsumersFromRecentJoinedConsumers(), which dropped an entry once themark-delete position passed the join position. The name conveys neither the key nor the value.
PersistentSubscription, fedexclusively by
StickyKeyDispatcher.getRecentlyJoinedConsumers(), whose only overrider was theclassic dispatcher. PIP-282 renamed the consumer-level sibling (
readPositionWhenJoining→lastSentPositionWhenJoining) but left this one alone, and PIP-379 then declared it removed.The equivalent modern observability is richer:
drainingHashes,drainingHashesCount,drainingHashesClearedTotal,drainingHashesUnackedMessagesandkeyHashRangeArrays.Verifying this change
This change is already covered by existing tests.
The end-to-end proof that the removal is behaviourally inert on the default path is
ConsumerStatsTest.testConsumerStatsOutput, which asserts the exact field set of the/admin/v2/.../statsJSON: after commit 2 it pins the modern Key_Shared stats shape, and aftercommit 5 it proves nothing else shifted.
Run locally, all green:
./gradlew quickCheck sanityCheck verifyTestGroups:pulsar-broker:test --tests "org.apache.pulsar.client.api.KeySharedSubscription*" --tests "org.apache.pulsar.client.impl.KeySharedSubscriptionMaxUnackedMessagesTest":pulsar-broker:test --tests "org.apache.pulsar.broker.service.persistent.*":pulsar-broker:test --tests "org.apache.pulsar.broker.stats.ConsumerStatsTest" --tests "org.apache.pulsar.broker.stats.AuthenticatedConsumerStatsTest" --tests "org.apache.pulsar.broker.admin.*TopicStats*":pulsar-broker:test --tests MessageRedeliveryControllerTest --tests TransactionConsumeTest:pulsar-common:testTwo places where coverage changes meaning rather than disappearing:
KeySharedSubscriptionTest.testContinueDispatchMessagesWhenMessageTTLhad a&& !impl.classicguard that suppressed the consumer3 assertion under the classic path. It is now unconditional,
matching the already-unconditional consumer2 check above it — the assertion is strengthened, not
dropped.
MessageRedeliveryControllerTest.testContainsStickyKeyHashesis rewritten onto the singularcontainsStickyKeyHash(int). The plural overload was an any-match loop, socontainsStickyKeyHashes(Set.of(104, 105))passed on 104 alone; the singular form now asserts 104present and 105 absent.
testRecentJoinedPosWillNotStuckOtherConsumeris kept — it guards#7105 and is implementation-agnostic — with its
comments reworded into draining-hashes terms.
Does this pull request potentially affect one of the following parts:
Public API:
ConsumerStats.getReadPositionWhenJoining()andgetKeyHashRanges()(both@Deprecatedsince 4.0.0) andSubscriptionStats.getConsumersAfterMarkDeletePosition()areremoved, along with the corresponding
ConsumerStatsImpl/SubscriptionStatsImplfields.Configuration:
subscriptionKeySharedUseClassicPersistentImplementationandsubscriptionSharedUseClassicPersistentImplementationare removed. Both defaulted tofalseandneither appeared in
conf/broker.conforconf/standalone.conf; a stale value is ignored ratherthan rejected (see Upgrade impact).
REST endpoints: no endpoint changes; the topic-stats response shape loses
consumersAfterMarkDeletePosition(see Upgrade impact).Documentation
doc-not-needed— the removed configuration keys were never documented inconf/broker.conf/conf/standalone.conf.