Cache store: SSD wear protection -- compaction policy, daily write cap, drive-type behaviour, write stats

sky_line commented

2026-10-12 11:17:22 +00:00

Owner

Part of #715 (cache rewrite epic).

Compaction adds writes; keep SSD wear at or below today's file-per-item cache.

Scale: a heavy explorer might write ~12 GB/day of new cache data: ~0.7%/year of a 600 TBW 1 TB TLC drive, ~3%/year of a 150 TBW 500 GB QLC drive. Small, but bounded anyway:

No compaction while the budget has room. Dead records stay until space is needed.
Reclaim only mostly-dead segments (<= 25% live), rescuing only records with a real reuse count; everything else is evicted with the segment. Worst-case write amplification 1.33x.
Daily compaction cap (e.g. 10% of the day's new data). Past the cap, whole segments are evicted without rescue -- extra misses, never extra wear.
No writes on reads: access times live in the index and are saved with its snapshot (today both caches rewrite file timestamps on access).
Drive-type behaviour: detect rotational vs. solid-state (Linux /sys/block/*/queue/rotational, Windows storage seek-penalty query). SSD: minimal compaction only, no locality rewrites. HDD: locality rewrites allowed (see the HDD locality ticket).
Stats: bytes written per day (new / compaction), shown in the cache stats and logged.

Large sequential appends and whole-segment deletes (one TRIM each) are also the SSD-friendliest write pattern, lowering the drive's own internal write amplification compared with small-file churn.

Part of #715 (cache rewrite epic). Compaction adds writes; keep SSD wear at or below today's file-per-item cache. **Scale:** a heavy explorer might write ~12 GB/day of new cache data: ~0.7%/year of a 600 TBW 1 TB TLC drive, ~3%/year of a 150 TBW 500 GB QLC drive. Small, but bounded anyway: 1. **No compaction while the budget has room.** Dead records stay until space is needed. 2. **Reclaim only mostly-dead segments** (<= 25% live), rescuing only records with a real reuse count; everything else is evicted with the segment. Worst-case write amplification 1.33x. 3. **Daily compaction cap** (e.g. 10% of the day's new data). Past the cap, whole segments are evicted without rescue -- extra misses, never extra wear. 4. **No writes on reads:** access times live in the index and are saved with its snapshot (today both caches rewrite file timestamps on access). 5. **Drive-type behaviour:** detect rotational vs. solid-state (Linux `/sys/block/*/queue/rotational`, Windows storage seek-penalty query). SSD: minimal compaction only, no locality rewrites. HDD: locality rewrites allowed (see the HDD locality ticket). 6. **Stats:** bytes written per day (new / compaction), shown in the cache stats and logged. Large sequential appends and whole-segment deletes (one TRIM each) are also the SSD-friendliest write pattern, lowering the drive's own internal write amplification compared with small-file churn.

sky_line added the

triage

perf

feature

labels

2026-10-12 11:17:22 +00:00

sky_line added this to the Slipstream project

2026-10-12 11:17:22 +00:00

sky_line referenced this issue

2026-10-12 11:17:23 +00:00

[Epic] Cache rewrite: segment store, compression, tiered texture caching, disk-aware sizing #715

sky_line commented

2026-10-18 22:52:43 +00:00

Author

Owner

Interface audit (step 1 of #715): wear-relevant writes today and this ticket's slice of the store.

Writes today that the store removes or bounds

Writer	Code	What it writes
Asset cache access time	llfilesystem.cpp:96-104, :380	file mtime on READ open, at most hourly per file
GPU-ready access time	llgputexcache.cpp:492 `touch`	mtime, hourly per file
Sound / cluster access time	llsoundcache.cpp:322, llcleanclustercache.cpp:420	mtime, hourly per file
J2C access time	lltexturecache.cpp:1387 + `writeUpdatedEntries` every 300 s (:908)	28-byte entries rewritten in place
J2C refinement	lltexturecache.cpp:546 `doWrite`	the whole body file again on every discard step
Asset purge	lldiskcache.cpp:92, every 15 min	deletes; stats 265k files per run

Slice of the interface

struct WearPolicy {
    bool ssd;                        // drive detection (Linux /sys/block/*/queue/rotational, Windows seek penalty)
    F32  reclaim_live_max = 0.25f;   // reclaim only segments <= 25% live
    U64  daily_compaction_cap_bytes; // 0 = 10% of the day's new bytes
    bool locality_rewrites;          // HDD only (#712)
};
Store::setWearPolicy(const WearPolicy&);
Store::touch(key);   // RAM only: last_access, prev_session_access, reuse_count, saved with the index snapshot
Stats: bytes_written_today_new, bytes_written_today_compaction, dead_bytes, segments

touch() never writes; access data rides the index snapshot (one sequential file).
Rescue during compaction is limited to records whose reuse_count > 0; everything else is evicted with the segment.
J2C upgrade appends count as new bytes (they replace today's body rewrites one for one). The writer queue's SUPERSEDE_PENDING flag drops an upgrade that is replaced before it reaches disk, which saves bytes compared with today.
Past the daily cap, whole segments are evicted without rescue.

Drop-in plan

Lands with the store core: in phase 1 only "no compaction while under budget" + whole-segment eviction + stats. Rescue, the daily cap and drive detection follow once eviction runs on real data. Stats show in the existing cache stats line and the log.

**Interface audit (step 1 of #715): wear-relevant writes today and this ticket's slice of the store.** ### Writes today that the store removes or bounds | Writer | Code | What it writes | |---|---|---| | Asset cache access time | llfilesystem.cpp:96-104, :380 | file mtime on READ open, at most hourly per file | | GPU-ready access time | llgputexcache.cpp:492 `touch` | mtime, hourly per file | | Sound / cluster access time | llsoundcache.cpp:322, llcleanclustercache.cpp:420 | mtime, hourly per file | | J2C access time | lltexturecache.cpp:1387 + `writeUpdatedEntries` every 300 s (:908) | 28-byte entries rewritten in place | | J2C refinement | lltexturecache.cpp:546 `doWrite` | the whole body file again on every discard step | | Asset purge | lldiskcache.cpp:92, every 15 min | deletes; stats 265k files per run | ### Slice of the interface ```cpp struct WearPolicy { bool ssd; // drive detection (Linux /sys/block/*/queue/rotational, Windows seek penalty) F32 reclaim_live_max = 0.25f; // reclaim only segments <= 25% live U64 daily_compaction_cap_bytes; // 0 = 10% of the day's new bytes bool locality_rewrites; // HDD only (#712) }; Store::setWearPolicy(const WearPolicy&); Store::touch(key); // RAM only: last_access, prev_session_access, reuse_count, saved with the index snapshot Stats: bytes_written_today_new, bytes_written_today_compaction, dead_bytes, segments ``` - `touch()` never writes; access data rides the index snapshot (one sequential file). - Rescue during compaction is limited to records whose `reuse_count` > 0; everything else is evicted with the segment. - J2C upgrade appends count as **new** bytes (they replace today's body rewrites one for one). The writer queue's SUPERSEDE_PENDING flag drops an upgrade that is replaced before it reaches disk, which saves bytes compared with today. - Past the daily cap, whole segments are evicted without rescue. ### Drop-in plan Lands with the store core: in phase 1 only "no compaction while under budget" + whole-segment eviction + stats. Rescue, the daily cap and drive detection follow once eviction runs on real data. Stats show in the existing cache stats line and the log.

sky_line referenced this issue

2026-10-18 22:52:43 +00:00

Cache store: hard-disk locality -- region-ordered writes and hot-set rewrites #712

sky_line referenced this issue

2026-10-18 23:31:24 +00:00

Cache store: hard-disk locality -- region-ordered writes and hot-set rewrites #712

sky_line referenced this issue

2026-10-18 23:31:24 +00:00

Texture disk tiers: one budget, GPU-ready cache as a reuse-admitted hot tier, caching strategy option, disk-space-based default sizes #705

sky_line referenced this issue

2026-10-19 00:25:41 +00:00

[Epic] Cache rewrite: segment store, compression, tiered texture caching, disk-aware sizing #715