[thweetkomputer]

Context

Deferred mimalloc purging can accumulate reclamation work that later executes on a request worker's allocation path. This changes the online default to immediate purging so eligible memory is reclaimed without the default 60-second delay.

Behavior before and after

• Default --mimalloc-purge-delay-ms: 60000 → 0.
• Recovery still uses mimalloc's native delay and completes the existing per-worker collection barriers before applying the online setting.
• Explicit CLI overrides remain supported, including 60000 to restore the previous policy and -1 to disable purging.

Implementation

Change kDefaultMimallocPurgeDelayMs, which initializes ServerOptions and the CLI's displayed default. Clarify zero in CLI help and document the reclamation tradeoff beside the constant.

Design decisions and alternatives

Immediate purge retains memory reclamation, unlike disabling purge. It shifts work toward frees; it does not make large frees cheap or eliminate ScanHashMap expansion latency.

Prior diagnostic measurements, not tests of this patch:

• In the aged-allocation experiment, the previous policy produced later malloc stalls of approximately 81–82 ms. With purge_delay=0, later malloc maxima were 0.274/0.276 ms, while freeing the touched 1 GiB still cost approximately 100 ms in total.
• A separate ScanHashMap insertion experiment with purge_delay=0 still observed approximately 10 ms when retiring a 128 MiB old table, including approximately 9.5 ms in MADV_DONTNEED. These are different experiments and must not be interpreted as an 82 ms → 10 ms map A/B comparison.

The existing SPDK server comparison used PR #366 at db2b5a5d42a3cbc0980b4c97b8905d94e07a940e: retain 1 billion keys, 80 connections issuing random overwrite SETs, 16 threads, 1024-byte values, no key prefix, and defrag block sleep 100 ms. Both data nodes were restarted for each arm; each arm had 120 seconds of warmup and 600 seconds measured using merged complete 5-second HDR windows.

Online purge delay	SET/s	Pooled p99.99 (ms)	Worst 5-second p99.99 (ms)	Max (ms)
Previous default, first run	254,854	0.874	3.625	5.451
0 ms	229,185	0.890	3.857	6.835
Previous default, repeat	221,801	0.925	3.421	5.675

The first run had no active primary defrag; both subsequent runs did, with different reclamation intensity. The measurements show no clear overall tail improvement or regression, and do not establish performance equivalence or isolate throughput effects. This overwrite workload does not exercise map doubling. All measured windows were complete with no HDR drops/overflow, connection errors were zero, and both nodes retained exactly 1 billion keys.

Documentation and comments

Updated the nearby policy comment and CLI help. No architectural lifecycle, module boundary, or recovery invariant changes.

Test plan

• pre-commit run clang-format --files include/lavik/server.h app/lavik.cpp: passed with clang-format 23.1.1.
• git diff --cached --check: passed.
• pre-commit run --files include/lavik/server.h app/lavik.cpp: blocked while initializing the unrelated Prettier hook by npm EALLOWGIT; the applicable C++ hook passed separately.
• Full Linux build/runtime tests were not run locally on macOS; PR CI supplies those checks. The measurements above exercised the explicit setting on earlier builds, not this new commit.

Risk assessment

Large or batched frees can synchronously purge and still stall a worker. Immediate reclamation can also increase later page-fault/recommit work. This is a default tuning change, not a complete fix for long-tail latency.

Rollback plan

Set --mimalloc-purge-delay-ms=60000 or revert this commit.

Reviewer guide

Review the default and tradeoff comment in include/lavik/server.h, then the help text in app/lavik.cpp. The post-recovery option application in src/redis/server.cpp is unchanged.

Follow-up work

Reducing the remaining synchronous cost of retiring large ScanHashMap tables requires a separate change.

Summary by CodeRabbit

• Changed
  • The default memory purge delay is now 0 milliseconds, so eligible pages are purged immediately during normal operation.
  • During recovery, mimalloc’s native purge delay remains in effect until workers finish.
  • Updated command-line help clarifies that the online purge delay applies after recovery; 0 purges immediately, while -1 disables purging.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

• Configuration used: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: 873ad61d-2c0f-4381-8be3-dc833eeb6e91

📥 Commits

Reviewing files that changed from the base of the PR and between 91f0c89 and 0c481a2.

📒 Files selected for processing: 2

• app/lavik.cpp
• include/lavik/server.h

Included review availability: This review used your included allowance. Your plan provides up to 1 included review per hour; 0 remain after this review.

Walkthrough

The default mimalloc purge delay changes from 60 seconds to 0 milliseconds. The command-line help explains that the setting applies after recovery, that 0 purges immediately, and that -1 disables purging.

Changes

Mimalloc purge delay

Layer / File(s)	Summary
Default and command-line help include/lavik/server.h, app/lavik.cpp	The default changes to 0 milliseconds. Comments describe immediate online purging, possible synchronous costs for large frees, and delayed purging during recovery. The command-line help documents when the setting applies and the meanings of 0 and -1.

Priority: ⬇️ Low

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Bug fix

Suggested reviewers: liunyl

Merge Risk: ⚪ Minimal · up to 0c481

The change is mergeable with normal CI validation; no actionable defect was found in the reviewed scope.

Pre-merge checks: 4 · 1

Failed checks: 1 warning

Check name	Status	Explanation	Resolution
Docstring Coverage		Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 2 functions across 2 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

Passed checks: 4 passed

Check name	Status	Explanation
Title check		The title clearly and concisely describes the primary change: making online mimalloc purging immediate by default.
Description check		The description covers all required sections, explains the behavior change and design tradeoffs, documents comments, lists validation commands and results, and includes risk, rollback, reviewer, and f…
Linked Issues check		Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check		Check skipped because no linked issues were found for this pull request.

Full details: Description check

Explanation

The description covers all required sections, explains the behavior change and design tradeoffs, documents comments, lists validation commands and results, and includes risk, rollback, reviewer, and follow-up information.

• Fix all pre-merge checks with AI

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{A rabbit checks the purge delay,
Zero now clears pages right away.
After recovery, workers run,
Then purging starts when work is done.
Large frees may pause the rabbit’s feet,
While minus one leaves purging asleep.}

_{Comment @coderabbitai help to get the list of available commands.}