Replace SubscribeEvictions with IRecordTriggers#1712
Conversation
7531edc to
4311e07
Compare
There was a problem hiding this comment.
Pull request overview
This PR migrates Garnet heap-size accounting on page eviction from the SubscribeEvictions iterator/observer scan path to Tsavorite’s per-record IRecordTriggers.OnEvict callback, while also threading an EvictionSource (main log vs read cache) through the eviction APIs.
Changes:
- Introduces
EvictionSource { MainLog, ReadCache }and threads it throughIRecordTriggers.OnEvict,IStoreFunctions.OnEvict, andIAllocator.EvictRecordsInRange. - Adds
AllocatorSettings.IsReadCache/AllocatorBase.IsReadCacheto tag eviction callbacks with the correct source. - Updates Garnet
CacheSizeTrackerwiring to useSetLogSizeTracker(fast-path size tracking) and moves eviction decrements toGarnetRecordTriggers.OnEvict.
Reviewed changes
Copilot reviewed 15 out of 15 changed files in this pull request and generated 1 comment.
Show a summary per file
| File | Description |
|---|---|
| libs/storage/Tsavorite/cs/src/core/Index/Tsavorite/Tsavorite.cs | Marks the read-cache allocator via AllocatorSettings.IsReadCache = true. |
| libs/storage/Tsavorite/cs/src/core/Index/StoreFunctions/StoreFunctions.cs | Updates store-functions eviction forwarding to include EvictionSource. |
| libs/storage/Tsavorite/cs/src/core/Index/StoreFunctions/IStoreFunctions.cs | Updates OnEvict signature to include EvictionSource. |
| libs/storage/Tsavorite/cs/src/core/Index/StoreFunctions/IRecordTriggers.cs | Adds EvictionSource enum and updates IRecordTriggers.OnEvict signature/docs. |
| libs/storage/Tsavorite/cs/src/core/Allocator/TsavoriteLogAllocatorImpl.cs | Switches allocator base ctor usage to pass AllocatorSettings directly. |
| libs/storage/Tsavorite/cs/src/core/Allocator/TsavoriteLogAllocator.cs | Updates EvictRecordsInRange signature to include EvictionSource. |
| libs/storage/Tsavorite/cs/src/core/Allocator/SpanByteAllocatorImpl.cs | Switches allocator base ctor usage to pass AllocatorSettings directly. |
| libs/storage/Tsavorite/cs/src/core/Allocator/SpanByteAllocator.cs | Updates EvictRecordsInRange signature to include EvictionSource. |
| libs/storage/Tsavorite/cs/src/core/Allocator/ObjectAllocatorImpl.cs | Threads EvictionSource into per-record eviction callbacks to storeFunctions.OnEvict. |
| libs/storage/Tsavorite/cs/src/core/Allocator/ObjectAllocator.cs | Updates wrapper to pass EvictionSource through EvictRecordsInRange. |
| libs/storage/Tsavorite/cs/src/core/Allocator/IAllocator.cs | Updates EvictRecordsInRange contract to include EvictionSource. |
| libs/storage/Tsavorite/cs/src/core/Allocator/AllocatorSettings.cs | Adds IsReadCache flag for allocator tagging. |
| libs/storage/Tsavorite/cs/src/core/Allocator/AllocatorBase.cs | Stores IsReadCache, passes EvictionSource into eviction calls, and updates recovery eviction path. |
| libs/server/Storage/SizeTracker/CacheSizeTracker.cs | Replaces SubscribeEvictions with SetLogSizeTracker for internal size tracking. |
| libs/server/Storage/Functions/GarnetRecordTriggers.cs | Enables CallOnEvict and implements OnEvict(ref LogRecord, EvictionSource) heap-size decrement routing. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| // Drives per-record heap-size decrement on page eviction. Mirrors the work the | ||
| // legacy SubscribeEvictions → LogSizeTracker.OnNext observer path used to perform | ||
| // (see CacheSizeTracker.Initialize for the wiring change). | ||
| public bool CallOnEvict => cacheSizeTracker != null; |
There was a problem hiding this comment.
CallOnEvict is currently gated on cacheSizeTracker != null, but GarnetServer.CreateStore always constructs a CacheSizeTracker and passes it into GarnetRecordTriggers even when heap-size tracking is disabled (only sets sizeTracker = cacheSizeTracker when LogMemorySize/ReadCacheMemorySize > 0). This makes CallOnEvict effectively always true and will invoke OnEvict on every page eviction, doing MemoryUtils.CalculateHeapMemorySize work even when there is no tracker initialized (and the decrement becomes a no-op). Consider gating on cacheSizeTracker.IsInitialized (or mainLogTracker/readCacheTracker != null) so eviction callbacks are only enabled when size tracking is actually configured.
7c62bf1 to
695d6ab
Compare
… tracking Garnet's `CacheSizeTracker` used `LogAccessor.SubscribeEvictions` to drive a `LogSizeTracker.OnNext` observer that, on every page eviction, allocated a scan iterator and walked the page to sum `MemoryUtils.CalculateHeapMemorySize` over each non-null / non-closed record. This is heavyweight for a hot path (buffer-pool allocation, iterator bookkeeping, epoch resume/suspend, and a virtual dispatch per page). This PR migrates the per-page heap-size decrement onto the per-record `IRecordTriggers.OnEvict` hook introduced in #1695, which the object allocator already walks during `EvictRecordsInRange`. That collapses the "scan iterator + observer + sum" path into a single per-record callback directly on the record we would have visited anyway. ### Tsavorite changes - Add `EvictionSource { MainLog, ReadCache }` and thread it through `IRecordTriggers.OnEvict`, `IStoreFunctions.OnEvict`, and `IAllocator.EvictRecordsInRange`. Garnet uses this to route decrements to the correct counter (`AddHeapSize` vs `AddReadCacheHeapSize`). - Change `IRecordTriggers.CallOnEvict` / `IStoreFunctions.CallOnEvict` from a property to `CallOnEvict(EvictionSource)` so the allocator can skip the per-record eviction walk entirely when the application has no work on that side (e.g. a read-cache-only budget should not force walks of the main-log allocator). - Add `AllocatorSettings.IsReadCache` and `AllocatorBase.IsReadCache`, set to `true` by `Tsavorite.cs` when constructing the read-cache allocator. This is the cleanest way to distinguish the two allocators at `OnPagesClosedWorker` time without relying on the `evictCallback` sentinel, and is used to pass the correct `EvictionSource` to `CallOnEvict`. - `AllocatorBase.EvictPageForRecovery` routes through the per-record `EvictRecordsInRange` when `storeFunctions.CallOnEvict(source)` is set. The legacy `MemoryPageScan(observer)` path is preserved as a fallback for consumers that still use `SubscribeEvictions`. - Collapse `AllocatorBase`'s constructor to accept `AllocatorSettings` directly instead of unpacking individual fields at each concrete allocator (`ObjectAllocatorImpl`, `SpanByteAllocatorImpl`, `TsavoriteLogAllocatorImpl`). - Tighten `ObjectAllocatorImpl.EvictRecordsInRange` to match `ObjectScanIterator`'s single-page invariant: clip `stopAddress` to the start page, bail on `offset == 0`, and document that both callers (`OnPagesClosedWorker`, `EvictPageForRecovery`) hand single-page ranges. ### Garnet changes - `GarnetRecordTriggers.CallOnEvict(EvictionSource)` returns true only for the sides that are actually configured (`mainLogTracker` and/or `readCacheTracker` non-null), avoiding pointless per-record walks on the untracked allocator. - `GarnetRecordTriggers.OnEvict(ref LogRecord, EvictionSource)` computes `MemoryUtils.CalculateHeapMemorySize(in logRecord)` and dispatches to `AddHeapSize(-size)` or `AddReadCacheHeapSize(-size)` based on the source. This goes through the standard asserting path — the counter must never undershoot zero. - `CacheSizeTracker.Initialize` replaces the two `SubscribeEvictions` calls with `SetLogSizeTracker` calls so the fast-path size tracking during `TryCopyToTail`, `TryCopyToReadCache`, and object-page growth (`UpdateSize`, `IncrementSize`) continues to work unchanged. ### MainStore heap-tracking fix (root cause of negative counter) Routing the eviction decrement through the asserting `AddHeapSize` path surfaced a pre-existing gap: `MainStore` record-creation paths never emitted a positive heap-size bump when a record's key or value spilled to overflow (large inline field backed by a heap-allocated byte array). The legacy observer path masked this silently because `OnNext` did a raw decrement with no assertion — the counter quietly went negative on every HLL sparse→dense transition (≈-12 KB per key) and accumulated drift elsewhere. `MainStore` now emits balanced heap accounting at all create/update/delete entry points that go through a typed function callback rather than the shared `SessionFunctionsUtils` writer (which already tracks deltas): - `RMWMethods.PostInitialUpdater`: `+logRecord.CalculateHeapMemorySize()` for the freshly-created record. - `RMWMethods.PostCopyUpdater`: `+dstLogRecord.CalculateHeapMemorySize()` for the new record only. The source is not subtracted — the `TSourceLogRecord` may be an in-memory main-log record (whose heap is tracked in `mainLogTracker` and will leak upward by a bounded sealed- source amount, parity with `ObjectStore`'s `ClearSource=false` branch and the legacy observer path), a read-cache record (heap lives in `readCacheTracker`, unrelated to this write), or a pending-IO `DiskLogRecord` (heap never counted in any tracker). Subtracting unconditionally would undercount in the last two cases and drive the counter negative. - `RMWMethods.InPlaceUpdater`: `GetValueHeapMemorySize()` pre/post delta on `Succeeded`, so value-heap changes (e.g. APPEND, SETRANGE, or any path that triggers `ReallocateValueOverflow` / `ConvertInlineToOverflow` / `ConvertOverflowToInline`) are tracked. - `UpsertMethods.PostInitialWriter` (all three overloads): `+logRecord.CalculateHeapMemorySize()` for SET-style inserts that create a new record through the Upsert path. - `DeleteMethods.InPlaceDeleter`: `-logRecord.CalculateHeapMemorySize()` before the wrapper sets Tombstone. After tombstone is set, `CalculateHeapMemorySize` short- circuits to zero and `EvictRecordsInRange` skips the record, so without this decrement the creation-side increments would leak for every DEL of an overflow record. The `HyperLogLogPFADD_LTM` suite is the regression that first surfaced this — HLL sparse→dense goes through CopyUpdater and allocates ≈12 KB of overflow for the dense representation. With the fix the counter stays balanced and the assertion `heapSize.Total >= 0` holds throughout. ### Parity Behavior at every record state encountered during page eviction is bit-for-bit identical to the prior `SubscribeEvictions` path: | Record state | Old path (iterator) | New path (per-record) | Delta | | ------------------------------------------------ | ---------------------------------- | ---------------------------- | ----- | | Valid, !Sealed, !Tombstone, !IsNull | yielded → -CalculateHeapMemorySize | OnEvict → -Calculate... | same | | `IsNull` | skipped by iterator | skipped by filter | 0 | | `SkipOnScan` (Invalid or Sealed) | skipped by iterator | skipped by filter | 0 | | `Tombstone` (post-delete, already ValueIsInline) | yielded → 0 via `!Info.Tombstone` | skipped by filter → 0 | 0 | ### Hot-path cost Fast paths (SET / GET / INCR on inline-sized values) pay only a few aggressive-inlined inline-bit checks; `GetValueHeapMemorySize` on an inline record early-returns 0 before any tracker call, and a `heap != 0` guard skips the `AddHeapSize` dispatch entirely. Tracker work occurs only when a record genuinely created, resized, or freed an overflow/object allocation — which was already a heavyweight event. ### Testing All on net10.0 Debug: - `HyperLogLogTests` (incl. `HyperLogLogPFADD_LTM{32,4096}`, `HyperLogLogTestPFMERGE_LTM_*`) ✓ - `CacheSizeTrackerTests` ✓ - `RespListTests` (incl. `ListPushPopStressTest` ×10), `RespHashTests`, `RespSetTests`, `RespSortedSetTests`, `RespEtagTests`, `RespBitmapTests`, `RespTests.Set*`, `RespTests.Del*`: 517/517 ✓ - Clean build on both `Garnet.slnx` and `Tsavorite.slnx` (0 warnings, 0 errors), `dotnet format --verify-no-changes` clean. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
695d6ab to
9e7454c
Compare
…yUpdater For in-memory CopyUpdate, Tsavorite's HeapObjectBase.CacheSerializedObjectData cloned src.ValueObject into dst.ValueObject and Garnet's PostCopyUpdater then cloned again, overwriting the first clone. Every ObjectStore CopyUpdate was paying for two Clone() calls and one immediate GC. Fix by making CacheSerializedObjectData the single point of cloning for all sources (memory and pending-IO DiskLogRecord), and stripping the redundant clone in Garnet's PostCopyUpdater. Tsavorite changes (IHeapObject / HeapObjectBase / InternalRMW): * Drop the unused 'ref LogRecord srcLogRecord' parameter from CacheSerializedObjectData; its only uses were an identity assert and a 'srcLogRecord.ValueObject'-as-'this' redirect. * Add an explicit 'bool srcIsOnMemoryLog' parameter so the method can safely run for DiskLogRecord sources (which cannot be expressed as ref LogRecord). When false, perform the clone and return immediately - skipping the serialization state machine, which is meaningless for ephemeral disk sources (the (v) data is already persisted on disk from a prior flush, 'this' is about to be disposed up the pending chain, and ClearSourceValueObject is ignored by InternalRMW for non-memory sources). * In InternalRMW, drop the 'isMemoryLogRecord' guard around the call, so cloning is always delegated to CacheSerializedObjectData regardless of source kind. Garnet changes (ObjectStore RMWMethods): * PostCopyUpdater now reads 'dstLogRecord.ValueObject' directly - the clone is always performed upstream. No conditional clone path remains here. The single surviving clone site still runs after the successful CAS, preserving CopyUpdater's deferred-allocation-until-CAS invariant. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2da959d to
1717f9b
Compare
…sal sites Before: ObjectStore PostCopyUpdater conditionally emitted +new-old when rmwInfo.ClearSourceValueObject was true, otherwise +new. This coupled PCU to a Tsavorite-internal signal and combined two concerns (creation of (v+1) and removal of (v)) into one branching call site. After: mirror MainStore's pattern by splitting into two unconditional sites. * PostCopyUpdater always emits +value.HeapMemorySize for the (v+1) creation. Matches PostInitialUpdater / PostInitialWriter already doing +new.heap. * OnDisposeValueObject(DisposeReason.CopyUpdated) emits -valueObject.HeapMemorySize for the (v) removal. This callback fires from InternalRMW.ClearValueIfHeap exactly when the source is cleared eagerly (ClearSourceValueObject=true and isMemoryLogRecord), which is the literal 'source freed now' signal. Checkpoint/disk paths that leave the source alive do not reach the dispose site; their decrement is emitted later by OnEvict when the sealed source page evicts (via CalculateHeapMemorySize, which includes ValueObject.HeapMemorySize). Net tracker state is bit-for-bit identical to the previous conditional form, but each site now has a single arithmetic direction and no awareness of Tsavorite's internal ClearSourceValueObject flag. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Simplify the record-lifecycle contract for IHeapObject values: - Remove IRecordTriggers.OnDisposeValueObject; OnDispose(ref LogRecord, reason) is now the single app-facing hook. Garnet's impl merges the Deleted and CopyUpdated branches into one heap-size decrement. - Drop the Action<IHeapObject> disposer lambda threaded through ObjectIdMap.Free, LogField.ClearObjectIdAndConvertToInline, LogRecord.ClearValueIfHeap/ClearHeapFields/Dispose, DiskLogRecord ctors, PendingContext.CopyFrom, ConditionalCopyToTail, scan iterators, and ISourceLogRecord.ClearValueIfHeap. - ObjectIdMap.Free(int) is now strictly slot reclamation; it never disposes. - DiskLogRecord.Dispose() calls ValueObject?.Dispose() itself, covering all its owners (PendingContext, AsyncIOContext, scan iterators, cluster migration / replication streaming, deserialized-from-disk). - InternalRMW CopyUpdated path routes through storeFunctions.OnDispose(ref srcMemLogRecord, CopyUpdated) followed by srcMemLogRecord.ClearValueIfHeap(). - ObjectAllocatorImpl.OnDispose(ref DiskLogRecord) becomes a no-op. - Update DeleteDisposeTests to the new single-hook surface. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Eliminates the asymmetry where DiskLogRecord.Dispose() unconditionally called ValueObject.Dispose() while on-log LogRecord disposal routed through the IRecordTriggers.OnDispose trigger. All record disposal now goes through IRecordTriggers, giving Garnet full control over whether to dispose the value object. - Add IRecordTriggers.OnDisposeDiskRecord(ref DiskLogRecord, DisposeReason) as a required interface member with explicit no-op overrides in DefaultRecordTriggers, SpanByteRecordTriggers, and GarnetRecordTriggers (avoids DIM/interface-dispatch overhead via JIT monomorphization). - Forward through IStoreFunctions/StoreFunctions to IAllocator, renamed IAllocator.OnDispose(ref DiskLogRecord) -> OnDisposeDiskRecord. ObjectAllocatorImpl forwards to storeFunctions (was no-op); SpanByteAllocator stays no-op (no IHeapObject); TsavoriteLogAllocator throws NotImplementedException. - Remove ValueObject?.Dispose() from DiskLogRecord.Dispose() — callers now fire the trigger before Dispose. Fixes latent bug in scan-iterator DiskLogRecord wrappers that shared ValueObject references with the still-alive on-log record. - Fire the trigger at all DiskLogRecord disposal sites: AllocatorBase (happy-path + retry/catch), AllocatorScan.GetFromDiskAndPushToReader, TsavoriteThread.InternalCompletePendingRequestFromContext, and the four cluster migrate/replication command sites. - Update test IRecordTriggers impls (TrackingRecordTriggers, ObjTrackingRecordTriggers) to satisfy the new interface member. All 102 relevant Tsavorite tests and 472 Garnet object-store tests pass. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…pose accounting Adds 12 tests that precisely assert when and how many times each IRecordTriggers callback fires across the full record lifecycle. Complements the existing DeleteDisposeTests with coverage of the new OnDisposeDiskRecord hook and the refactored DisposeReason / EvictionSource discriminators introduced in this PR. Trackers: - TrackedObjectValue (subclass of TestObjectValue) counts per-instance IHeapObject.Dispose invocations so tests can catch double-dispose across any combination of on-log, scan, pending-read, and eviction paths. - LifecycleTracker counts OnDispose/OnDisposeDiskRecord per DisposeReason, OnEvict per EvictionSource, plus OnFlush / OnDiskRead — with toggleable CallOn* flags so tests can verify gating. Covered behaviours: - OnDispose(CopyUpdated) fires exactly once on immutable RMW; value dispose count is exactly 1 when handler opts in. - OnDispose(Deleted) fires exactly once on delete; no OnDisposeDiskRecord. - In-memory scan fires OnDisposeDiskRecord(DeserializedFromDisk) exactly N times for N records and does NOT dispose shared value objects. - Disk scan fires the same hook exactly K times and disposes exactly K values when opt-in, with no double-dispose. - Pending read from disk fires exactly one OnDisposeDiskRecord, no OnDispose, no OnEvict. - Page eviction fires OnEvict(MainLog) exactly (live - tombstoned) times with no re-firing of OnDispose(Deleted) for tombstones. - CallOnFlush / CallOnDiskRead / per-source CallOnEvict gating fully suppress their respective callbacks when false. - Read cache eviction fires OnEvict(ReadCache) and not OnEvict(MainLog). - No value object is ever disposed more than once across an upsert → delete → re-upsert → evict → pending-read cycle. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The earlier refactor in eeb6a3b split ObjectStore PCU heap-size tracking into two sites (PCU emits +new unconditionally, OnDispose(CopyUpdated) emits -old when the source is cleared eagerly). The same split was needed in UnifiedStore, but its PostCopyUpdater was missed. Result: on an EXPIRE/PERSIST that went through CopyUpdate for an object record (e.g. SADD followed by EXPIRE), UnifiedStore.PCU still emitted +new - old while GarnetRecordTriggers.OnDispose(CopyUpdated) also emitted -old, double-decrementing the tracked heap size. A subsequent OnDispose(Deleted) from ReinitializeExpiredRecord then drove the counter below zero and tripped the Debug.Assert in LogSizeTracker.IncrementSize (HeapSize.Total should be >= 0). Fix: mirror ObjectStore — PCU emits +new.HeapMemorySize unconditionally, and the matching -old is emitted at the removal site (OnDispose(CopyUpdated) or OnEvict for checkpoint/disk paths). Fixes the four CI failures surfaced on the replace-subscribe-eviction branch: - RespTests.ReAddExpiredKey - RespTransactionProcTests.TransactionObjectExpiryProcTest - RespCustomCommandTests.CustomObjectCommandTest2 - RespSortedSetGeoTests.CanUseGeoSearchStoreWithDeleteKeyWhenSourceNotFound Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…cord disposal Three fixes from independent design reviews: 1. GarnetRecordTriggers.OnDispose(Deleted): use CalculateHeapMemorySize instead of only ValueObject.HeapMemorySize. This ensures overflow key/value bytes on MainStore string records are correctly decremented on RMW expire paths (ExpireAndResume, ExpireAndStop) where InPlaceDeleter does not run. CalculateHeapMemorySize returns 0 for tombstoned records, so this is a natural no-op on the mutable Delete() path where MainStore.InPlaceDeleter already subtracted before tombstone was set. 2. Cluster migration/replication: prevent double-trigger of OnDisposeDiskRecord by resetting diskLogRecord to default after the normal-path dispose, and guarding the finally/catch block with IsSet. Previously, a non-noop trigger implementation could receive a callback on a default/already-disposed record. 3. IRecordTriggers.OnDisposeDiskRecord: make it a default interface method (no-op) instead of requiring explicit implementation. This reduces the implementation burden on consumers who don't need disk-record lifecycle hooks. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…heap leak EvictRecordsInRange previously skipped all SkipOnScan records (Sealed OR Invalid). This leaked tracked heap for sealed-but-valid source records from mutable-region CopyUpdates and immutable-region deletes, whose overflow key/value bytes were never decremented from the tracker. Fix: skip only Invalid records (already disposed/elided) and Tombstoned records (already decremented at delete site). Sealed-but-Valid records are now visited by OnEvict, which correctly picks up their remaining heap contribution (overflow bytes, and value objects kept alive during checkpoint). For records where OnDispose(Deleted) already cleared the value (immutable delete), CalculateHeapMemorySize returns only key overflow (if any) — no double-decrement because OnDispose decremented the full amount when the record was not yet tombstoned, and ClearHeapFields(clearKey=false) zeroed the value slot afterward. OnEvict sees only what remains (key overflow if present, typically 0 for inline keys). Updated PageEvictionFiresOnEvictForEveryLiveRecord test to use bounds assertions reflecting that sealed source records from immutable-region deletes are now correctly visited. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…e state Session functions (ISessionFunctions) should not be responsible for heap-size tracking on destruction paths — that's IRecordTriggers' job. Two places were setting Tombstone BEFORE OnDispose(Deleted) fired, causing CalculateHeapMemorySize to return 0 and preventing the trigger from decrementing the tracked heap: 1. SessionFunctionsWrapper.InPlaceDeleter: set Tombstone+Dirty before returning to InternalDelete, which then called OnDispose(Deleted). Fixed by moving SetTombstone+SetDirtyAndModified to InternalDelete after OnDispose. 2. SessionFunctionsWrapper.InPlaceUpdater (ExpireAndStop): set Tombstone before returning to InternalRMW, which then called OnDispose(Deleted). Fixed by moving SetTombstone+SetDirtyAndModified to InternalRMW after OnDispose. This removes the manual cacheSizeTracker decrement from MainStore.InPlaceDeleter (it was a workaround for the wrong ordering) and makes GarnetRecordTriggers.OnDispose the single destruction-side decrement path for all record types. Also tightened PageEvictionFiresOnEvictForEveryLiveRecord test to use precise counts by observing TailAddress movement to distinguish mutable vs immutable deletes. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
c6d6dd9 to
61089a1
Compare
The CopyUpdated path is a partial clear (value-object slot only, key stays on the record). Expecting IRecordTriggers implementers to know this nuance — that they should subtract only ValueObject.HeapMemorySize and not the full CalculateHeapMemorySize — is error-prone. Move the accounting entirely into Tsavorite's InternalRMW: - Decrement logSizeTracker by ValueObject.HeapMemorySize before clearing - Call IHeapObject.Dispose() on the freed value object - Then ClearValueIfHeap() nulls the ObjectIdMap slot Remove OnDispose(CopyUpdated) from the trigger call site — the trigger is no longer involved in CopyUpdate accounting. GarnetRecordTriggers.OnDispose now only handles DisposeReason.Deleted. The IRecordTriggers doc is updated to reflect that CopyUpdated is handled internally. The trigger contract is now simple: - OnDispose(Deleted): subtract CalculateHeapMemorySize (full record heap) - OnEvict: subtract CalculateHeapMemorySize (whatever remains at eviction) - Creation sites (PostInitialUpdater/Writer/CopyUpdater, InPlaceUpdater): add the new/changed heap Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
957c5c4 to
89cbce5
Compare
- EvictRecordsInRange: update XML doc and skip-condition comments to reflect that sealed records are now visited (not skipped) and InPlaceDeleter no longer does heap tracking - GarnetRecordTriggers.OnDispose: simplify comment now that tombstone is always set AFTER OnDispose in all paths - Remove legacy SubscribeEvictions reference from EvictRecordsInRange doc Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…w double-subtract OnDispose(Deleted) now uses GetValueHeapMemorySize() instead of CalculateHeapMemorySize(). This prevents double-subtracting key overflow bytes for immutable-region deletes where: 1. OnDispose(Deleted) fires on the source (subtracts full heap including key) 2. ClearHeapFields(clearKey=false) keeps key alive for chain traversal 3. Source is sealed (not tombstoned) — OnEvict visits it later 4. CalculateHeapMemorySize returns key overflow → subtracted again With the fix: - OnDispose(Deleted) subtracts value-only heap (value overflow + value object) - For mutable deletes: tombstoned → OnEvict skips → key overflow is a bounded phantom freed by GC when the page is freed (same as pre-IRecordTriggers) - For immutable deletes: sealed → OnEvict visits → subtracts key overflow once - Net: no double-decrement, bounded key-overflow phantom on mutable deletes Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
138cced to
53389b6
Compare
Heap-Size Accounting: Record Lifecycle TraceAll destruction-side heap accounting is internal to Tsavorite via Accounting ModelCreation (+value): Session functions (PostInitialUpdater/Writer, PostCopyUpdater, InPlaceUpdater) emit Destruction (−key, −value): Tsavorite handles all decrements internally:
Lifecycle Trace
|
53389b6 to
8e8a852
Compare
Tsavorite now handles ALL heap-size decrements internally via logSizeTracker: - OnDispose(Deleted): decrements value heap before ClearHeapFields - EvictRecordsInRange: decrements key overflow for ALL records (including tombstoned), decrements value heap for non-tombstoned records - CopyUpdated: decrements value-object heap (already internalized earlier) GarnetRecordTriggers.OnDispose and OnEvict are now no-ops for accounting. CallOnEvict returns false. The trigger callbacks remain available for app-level resource cleanup (e.g. IDisposable.Dispose on value objects that hold external resources), but Garnet does not use them. The heap-tracking contract is now cleanly split: - Session functions: emit +value at creation sites (only the app knows the heap size of a newly created value) - Tsavorite: emit -key and -value at all destruction/eviction sites (the record is in hand, Tsavorite can read the sizes directly) EvictRecordsInRange is now called whenever logSizeTracker is set OR CallOnEvict is true, ensuring internal accounting runs even when the application opts out of the OnEvict trigger. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
8e8a852 to
3b96c8c
Compare
Key overflow was only tracked (+key) for internal TryCopyToTail/TryCopyToReadCache copies via logSizeTracker.UpdateSize. Records created by session functions (RMW, Upsert, Delete) never added +key, but EvictRecordsInRange subtracted -key at eviction — causing the tracker to undercount for overflow-key records. Fix: add logSizeTracker.IncrementSize(+key) at the CAS success site in InternalRMW, InternalUpsert, and InternalDelete. This pairs with the -key emitted by EvictRecordsInRange for all records (including tombstoned). The heap-tracking responsibility is now cleanly split: - Tsavorite: all key accounting (+key at CAS, -key at eviction) and all destruction-side value accounting (-value at OnDispose/eviction/CopyUpdated) - Session functions: +value at creation sites only Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Previously, every ISessionFunctions implementation (MainStore, ObjectStore, UnifiedStore, VectorStore) had to manually call cacheSizeTracker.AddHeapSize in PostInitialUpdater, PostInitialWriter, and PostCopyUpdater. This was error-prone — VectorSessionFunctions was missing these calls entirely, causing heap tracking to go negative on delete (server crash). Move all creation-site +value tracking into Tsavorite's CAS success sites and ReinitializeExpiredRecord: - InternalRMW: +value after PostInitialUpdater and PostCopyUpdater - InternalUpsert: +value after PostInitialWriter - ReinitializeExpiredRecord: +value after PostInitialUpdater (IPU path) ISF implementations now only handle ±delta for in-place updates (InPlaceUpdater/InPlaceWriter), where only the app knows the before/after sizes. All other heap tracking is fully internal to Tsavorite. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Replace SubscribeEvictions with IRecordTriggers.OnEvict for heap-size tracking
Garnet's
CacheSizeTrackerusedLogAccessor.SubscribeEvictionsto drive aLogSizeTracker.OnNextobserver that, on every page eviction, allocated a scan iterator and walked the page to sumMemoryUtils.CalculateHeapMemorySizeover each non-null / non-closed record. This is heavyweight for a hot path (buffer-pool allocation, iterator bookkeeping, epoch resume/suspend, and a virtual dispatch per page).This PR migrates the per-page heap-size decrement onto the per-record
IRecordTriggers.OnEvicthook introduced in #1695, which the object allocator already walks duringEvictRecordsInRange. That lets us collapse the "scan iterator + observer + sum" path into a single per-record callback directly on the record we would have visited anyway.Tsavorite changes
EvictionSource { MainLog, ReadCache }and thread it throughIRecordTriggers.OnEvict,IStoreFunctions.OnEvict, andIAllocator.EvictRecordsInRange. Garnet uses this to route decrements to the correct counter (AddHeapSizevsAddReadCacheHeapSize).AllocatorSettings.IsReadCacheandAllocatorBase.IsReadCache, set totruebyTsavorite.cswhen constructing the read-cache allocator. This is the cleanest way to distinguish the two allocators atOnPagesClosedWorkertime without relying on theevictCallbacksentinel.AllocatorBase.EvictPageForRecoverynow also routes through the per-recordEvictRecordsInRangewhenstoreFunctions.CallOnEvictis set. The legacyMemoryPageScan(observer)path is preserved as a fallback for consumers that still useSubscribeEvictions.AllocatorBase's constructor to acceptAllocatorSettingsdirectly instead of unpacking individual fields at each concrete allocator (ObjectAllocatorImpl,SpanByteAllocatorImpl,TsavoriteLogAllocatorImpl).Garnet changes
GarnetRecordTriggers.CallOnEvictis now gated oncacheSizeTracker != null.OnEvict(ref LogRecord, EvictionSource)computesMemoryUtils.CalculateHeapMemorySize(in logRecord)and dispatches toAddHeapSize(-size)orAddReadCacheHeapSize(-size)based on the source.CacheSizeTracker.Initializereplaces the twoSubscribeEvictionscalls withSetLogSizeTrackercalls so the fast-path size tracking duringTryCopyToTail,TryCopyToReadCache, and object-page growth (UpdateSize,IncrementSize) continues to work unchanged.Parity
Behavior at every record state encountered during page eviction is bit-for-bit identical to the prior
SubscribeEvictionspath:IsNullSkipOnScan(Invalid or Sealed)Tombstone(post-delete, already ValueIsInline)!Info.TombstoneTombstones converge via two independent mechanisms (the old relies on the
if (!Info.Tombstone)guard insideCalculateHeapMemorySize; the new short-circuits in the filter), so both yield 0 contribution as required given the delete-siteOnDispose(Deleted)already decremented the tracker.Testing
All tests pass on net10.0 Debug:
CacheSizeTrackerTests2/2RespListTests.ListPushPopStressTest(10 repetitions) ✓RespListTests,RespHashTests,RespSetTests,RespSortedSetTests: 467/467DeleteDisposeTests+ReadCacheTests: 228/228, 14 skippedGarnet.slnxandTsavorite.slnx(0 warnings, 0 errors).Future TODO
The only residual subtlety worth flagging (and pre-existing to this PR) is that a
TryCopyToTailproduces two records sharing the sameIHeapObjectviaTryCopyFrom. If a subsequent in-place RMW on the dest mutates that shared object (growing from X to Y),CalculateHeapMemorySizeat source-evict time will return the current Y, not the originally-added X — so the source's eviction under-/over-decrement scales with post-copy mutation. The old observer-iterator path reads the same field at the same time, so it has exactly the same behavior.