Redis Fork and Copy-on-Write Latency Tuning Prompt

You are a senior SRE who tunes Redis persistence forking so background saves stop causing latency spikes.

I will provide:
- `INFO persistence` (rdb_last_fork_usec / latest_fork_usec, aof/rdb in-progress), `INFO stats`
- Host RAM, `used_memory`, and OS settings (THP, `vm.overcommit_memory`)
- When the spikes occur relative to BGSAVE / AOF rewrite

Your job:

1. **Explain the mechanism**: `BGSAVE` and AOF rewrite `fork()` a child that shares the parent's memory copy-on-write. The `fork()` itself pauses the primary while the kernel copies page tables — proportional to dataset size — reported as `latest_fork_usec`. Afterward, every page the primary writes must be COW-duplicated, transiently inflating RSS.
2. **Read the signals**: `INFO persistence` → `latest_fork_usec` (a spike here == the pause), `rdb_last_cow_size`/`aof_last_cow_size` (COW memory used), `rdb_bgsave_in_progress`; correlate with `redis-cli --latency` and app p99 sawtooth aligned to save times.
3. **Kill Transparent Huge Pages**: THP makes COW copy 2 MB pages instead of 4 KB, hugely amplifying fork COW latency and memory. Disabling THP (`madvise`/`never`) is Redis's standard recommendation — Redis even logs a warning when THP is enabled.
4. **Set overcommit**: `vm.overcommit_memory = 1` so `fork()`/`BGSAVE` doesn't fail with "Cannot allocate memory" when free RAM < `used_memory` even though COW means it usually won't be touched.
5. **Reduce fork frequency/cost**: relax `save` points, prefer AOF with `auto-aof-rewrite-percentage` tuned so rewrites are rarer, or offload persistence to a replica (persist there, keep the primary save-light). Smaller datasets fork faster — another reason to shard huge instances.
6. **Ensure headroom**: keep enough free RAM for the COW spike so the fork doesn't trigger the OOM killer mid-save.
7. **Validate**: after changes, confirm `latest_fork_usec` and the p99 spike both drop.

Mark DESTRUCTIVE or risky: disabling persistence entirely to hide the spike (data loss), running with no free-RAM headroom (fork OOM), and leaving THP enabled (well-known Redis latency amplifier).

---

INFO persistence + host RAM: [PASTE]
THP / overcommit settings: [PASTE]
Spike timing vs saves: [DESCRIBE]