Fix heap size double-subtraction causing server crash under low-memory object operations

[badrishc]

Problem

Under lowMemory: true settings (LogMemorySize="2k", PageSize="512") with concurrent object operations (LPUSH/RPOP), the server crashes with ConnectionReset — killing the client connection while commands are in-flight. The ListPushPopStressTest fails intermittently (~90% of the time in Debug mode).

Root Cause

CacheSizeTracker called SetLogSizeTracker() to register the LogSizeTracker, but this only sets the logSizeTracker field — it does not subscribe the tracker as onEvictionObserver. Without the eviction observer, the OnNext callback is never invoked when pages are evicted.

To compensate, ResizeIfNeeded pre-computed a heapTrimmedSize by scanning records before eviction, then subtracted it from heapSize after ShiftAddresses completed. This introduced a race condition: between the scan and the subtraction, concurrent CopyUpdate operations on other threads could clear the source record's value object (via ClearSourceValueObject) and decrement heapSize via AddHeapSize(sizeAdjustment). The resizer then subtracted heapTrimmedSize — which still included the already-decremented record — causing a double-subtraction that drove heapSize to -80, triggering DebugAssertException → session crash → ConnectionReset.

Fix

1. CacheSizeTracker.cs: Use SubscribeEvictions() instead of SetLogSizeTracker(). This sets both onEvictionObserver and logSizeTracker, enabling the OnNext callback to subtract each record's current HeapMemorySize at actual eviction time. Records already cleared by concurrent CopyUpdate have HeapMemorySize=0, so they correctly contribute nothing.

2. LogSizeTracker.cs: Remove the stale heapSize.Increment(-heapTrimmedSize) from ResizeIfNeeded. This is no longer needed because OnNext now handles heap subtraction during ShiftAddresses → OnPagesClosed → MemoryPageScan. The heapTrimmedSize from DetermineEvictionRange is still computed and used for the eviction range calculation (how far to advance HeadAddress), just not for the actual heap counter adjustment.

3. RespListTests.cs: Convert ListPushPopStressTest to async to avoid thread pool starvation from blocking pop retries.

Validation

• ListPushPopStressTest (100 keys × 500 items, 200 concurrent tasks, Debug mode): 10/10 passes with [Repeat(10)] = 100 total stress iterations (was crashing ~90%)
• Full test suite: 3359 total, 0 failed, 3326 passed, 33 skipped

[Copilot]

Pull request overview

Updates ListPushPopStressTest to be more resilient under thread pool starvation by avoiding blocking Redis calls and switching to an async wait pattern.

Changes:

• Convert ListPushPopStressTest to an async Task NUnit test.
• Replace synchronous ListRightPop/Task.WaitAll with ListRightPopAsync/Task.WhenAll.
• Replace tight Thread.Yield() spin with a small async delay between empty-pop retries.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.