Issue Summary

The victim cache (disaggregated storage block cache) stores compressed page images on eviction via __wt_blkcache_compress, but that compression work is currently untimed. We have no visibility into how long victim-cache compression takes, which is needed to evaluate the "acceptable overhead" leg of the disagg victim-cache decision rule. This ticket adds a connection-level latency histogram, perf_hist_disaggcompress_latency_*, measured at the eviction call site.

Context

• On eviction, pages destined for the disagg victim cache are compressed at src/evict/evict_page.c:141 via __wt_blkcache_compress (defined in src/block_cache/block_io.c:664). The call is wrapped in WT_IGNORE_RET.
• The compress path already has DSRC counters (compress_write, compress_write_fail, compress_write_too_small, and a compression-ratio histogram), but no timing data.
• __wt_blkcache_compress is shared with the normal block-manager write path (src/block_cache/block_io.c:795), so timing must be done at the eviction caller to isolate the victim-cache cost rather than inside the shared function.
• Existing model to mirror: the disagg block-manager read/write latency histograms (perf_hist_disaggbmread_latency_ / perf_hist_disaggbmwrite_latency_), which are disagg-scoped and bucketed in microseconds — the right scale for per-block compression.

Proposed Solution

A microsecond latency histogram via the existing WT_STAT_USECS_HIST_INCR_FUNC machinery. This emits a _total_usecs accumulator plus 8 distribution buckets in one helper, giving both the mean (total_usecs / count) and the tail.

1. dist/stat_data.py — declare the histogram family

Add 9 PerfHistStat entries in the conn_stats list, alphabetically right after the perf_hist_disaggbmwrite_latency_* block. Bucket suffixes must match the macro at src/include/stat.h:356 exactly:

PerfHistStat\('perf\_hist\_disaggcompress\_latency\_lt100',   'disagg victim cache compress latency histogram \(bucket 1\) \- 50\-99us'\),
PerfHistStat\('perf\_hist\_disaggcompress\_latency\_lt250',   'disagg victim cache compress latency histogram \(bucket 2\) \- 100\-249us'\),
PerfHistStat\('perf\_hist\_disaggcompress\_latency\_lt500',   'disagg victim cache compress latency histogram \(bucket 3\) \- 250\-499us'\),
PerfHistStat\('perf\_hist\_disaggcompress\_latency\_lt1000',  'disagg victim cache compress latency histogram \(bucket 4\) \- 500\-999us'\),
PerfHistStat\('perf\_hist\_disaggcompress\_latency\_lt2500',  'disagg victim cache compress latency histogram \(bucket 5\) \- 1000\-2499us'\),
PerfHistStat\('perf\_hist\_disaggcompress\_latency\_lt5000',  'disagg victim cache compress latency histogram \(bucket 6\) \- 2500\-4999us'\),
PerfHistStat\('perf\_hist\_disaggcompress\_latency\_lt10000', 'disagg victim cache compress latency histogram \(bucket 7\) \- 5000\-9999us'\),
PerfHistStat\('perf\_hist\_disaggcompress\_latency\_gt10000', 'disagg victim cache compress latency histogram \(bucket 8\) \- 10000us\+'\),
PerfHistStat\('perf\_hist\_disaggcompress\_latency\_total\_usecs', 'disagg victim cache compress latency histogram total \(usecs\)'\),

2. src/include/block_inline.h — instantiate the helper

Add next to the existing disagg instantiations (currently lines 20-21):

WT\_STAT\_USECS\_HIST\_INCR\_FUNC\(disaggcompress, perf\_hist\_disaggcompress\_latency\)

This generates __wt_stat_usecs_hist_incr_disaggcompress(session, usecs), visible in evict_page.c.

3. src/evict/evict_page.c — bracket the call

Gate the clock with the standard pattern from the write path (src/block_cache/block_io.c:841) so __wt_clock is not paid when stats are disabled:

/\* Optionally compress the data before caching. \*/
timer = WT\_STAT\_ENABLED\(session\) && \!F\_ISSET\(session, WT\_SESSION\_INTERNAL\);
time\_start = timer ? \_\_wt\_clock\(session\) : 0;
WT\_IGNORE\_RET\(
  \_\_wt\_blkcache\_compress\(session, &buf\_orig, false, &compressed\_buf, NULL, &compressed\)\);
if \(timer\)
    *wt\_stat\_usecs\_hist\_incr\_disaggcompress\(
      session, WT\_CLOCKDIFF\_US\(*wt\_clock\(session\), time\_start\)\);
if \(compressed\_buf \!= NULL\)
    cache\_buf = compressed\_buf;

Plus declare uint64_t time_start; and bool timer; in the function's declaration block.

4. dist/s_all — regenerate

Run cd src/third_party/wiredtiger/dist && ./s_all to regenerate the derived files from stat_data.py (src/include/stat.h struct fields, src/support/stat.c clear/aggregate/init). These regenerated files are committed but must never be hand-edited.

Definition of Done

• cd dist && ./s_all exits clean (stat ordering + codegen validated).
• Build compiles
• A disagg workload with the victim cache enabled shows perf_hist_disaggcompress_latency_total_usecs advancing alongside compress_write in the statistics output.

Notes

• These are connection-level stats (the macro uses WT_STAT_CONN_INCR_), consistent with the other perf_hist__ families — distinct from the existing DSRC compress_write* counters.
• Measuring at the caller captures total per-block cost (pre_size + alloc + codec + memcpy), which is what the overhead-acceptability decision wants. If pure codec time across all callers is preferred instead, the alternative is bracketing src/block_cache/block_io.c:721.
• Touches the eviction/storage hot path — worth a sys-perf patch before merge.