v0.3.13 — fix preload-time blowup in multitenant-tagged specs

Cuts preload from ~28 min → ~15 sec by switching 5 SET preloads from -c 50 -t 4 -n allkeys (200 clients × 10M keys = 2B ops) to -c 1 -t 1 (exactly 10M ops, 1 SET per key). Test phases unchanged. Merged in #388.

v0.3.12 — multi-tenant tagged-keyname specs at 10M scale

Adds 7 benchmark specs that exercise a long-keyname multi-tenant access pattern not previously covered. Each key uses a unique hash-tag prefix plus a ~78-char structured suffix:
`{T:N}_metric_a_window_b_period_c_segment_d_aggregate_e_v1_v2_v3_v4_v5_v6`

This stresses ~6× more bytes hashed/compared per dict lookup than the default short `memtier-N` prefix used by sister specs.

Specs added

• `memtier_benchmark-multiple-hll-pfadd-10Mkeys-10B-values` — PFADD on 10M sparse HLLs
• `memtier_benchmark-10Mkeys-string-mget-80keys-100B-multitenant-tagged` — 80-key MGET batches
• `memtier_benchmark-10Mkeys-generic-expire-multitenant-tagged` — EXPIRE 7200 on long-keyname keys
• `memtier_benchmark-10Mkeys-string-incrby-multitenant-tagged` — INCRBY 1 on int-encoded keys
• `memtier_benchmark-10Mkeys-string-incrby-1000-multitenant-tagged` — INCRBY larger value variant
• `memtier_benchmark-10Mkeys-mixed-incrby-expire-multitenant-tagged` — mixed write workload (INCRBY:EXPIRE = 9:20 via `--command-ratio`)
• `memtier_benchmark-multiple-hll-pfcount-1key-10Mkeys-multitenant-tagged` — single-key PFCOUNT

All target `oss-standalone`. Used-memory at full scale: PFADD ~1.7 GB, MGET ~2.3 GB, EXPIRE ~2.3 GB, INCRBY ~1.25 GB, mixed ~1.25 GB, PFCOUNT ~1.7 GB. All request 8 GB resource headroom.

Merged in #387.

v0.3.11

What's new

session-caching specs: replace Lua preload with memtier preload_tool + add pipeline=16 variant (#386)

• memtier_benchmark-playbook-session-caching-hash-100k-sessions.yml: drop the init_lua block and use dbconfig.preload_tool with memtier_benchmark (--random-data + --data-size-range=200-300). Aligns with the standard memtier-preload idiom used elsewhere in the suite. Adds check.keyspacelen: 100000 sanity gate.
• New memtier_benchmark-playbook-session-caching-hash-100k-sessions-pipeline-16.yml: same workload with client --pipeline=16 to amplify per-call overhead and exercise reply-buffer / output-list paths.

Full Changelog: v0.3.10...v0.3.11

v0.3.10

Add 3 string SET key value GET (no expire) memtier specs across EMBSTR + RAW value sizes:

• memtier_benchmark-10Mkeys-string-set-with-get-option-10B (EMBSTR, ≤ 44 B)
• memtier_benchmark-1Mkeys-string-set-with-get-option-100B (RAW)
• memtier_benchmark-1Mkeys-string-set-with-get-option-1KB (RAW)

All three exercise the propagation argv-strip path in setGenericCommand() (src/t_string.c), which existing string benchmarks (separate SET / GET) did not cover.

Merged via #385.

v0.3.8

What's new

Adds 4 stream GC coverage specs (PR #382) to validate streamEntryIsReferenced() and streamCleanupEntryCGroupRefs() at larger consumer-group counts and across delete-strategy variants:

Spec	CG count	Command mix
stream-500-cgroups-xadd-xtrim-aggressive-gc	500	70 % XADD + 30 % XTRIM MAXLEN ~ 1000
stream-5000-cgroups-xadd-xtrim-aggressive-gc	5,000	same
stream-50-cgroups-xadd-xreadgroup-xdel	50	XADD (30%) + XDEL deterministic IDs (50%) + XREADGROUP (20%)
stream-500-cgroups-xadd-xreadgroup-xdel	500	same

Complements the existing stream-50-cgroups-xadd-xtrim-* specs by scaling CG count (10× / 100×) and adding the XDEL path (middle-of-stream delete triggering streamCleanupEntryCGroupRefs(), distinct from XTRIM's tail-trim).

Full Changelog: v0.3.7...v0.3.8

v0.3.7

Hotfix

• Add missing clientconfig.resources to EMBSTR validation specs (#381) — the 3 specs added in 0.3.5/0.3.6 were missing resources: inside clientconfig, causing KeyError: 'resources' in the coordinator and aborting every run before benchmark data was collected.

Full Changelog: v0.3.6...v0.3.7

v0.3.6

Changes

• Add 10M-keys EMBSTR validation spec (#380) — DRAM-bound sibling of the 1M × 10B spec for robj-layout validation. At 10M keys, used_memory scales to ~660MB which no longer fits in L3 on any fleet CPU — measures the cache-miss cost of pointer chasing that the L3-resident 1M variant hides.

Full Changelog: v0.3.5...v0.3.6

v0.3.5

Changes

• Add two EMBSTR-layout validation specs (#379) — APPEND-path test (10B EMBSTR → 40B RAW conversion) and mixed EMBSTR/RAW GET pipeline-100 test. Both use --random-data. Used to validate Valkey-#2516-style robj layout changes.

Full Changelog: v0.3.4...v0.3.5

v0.3.4

Hotfix

• fix(coordinator): correct compute_regression_table args + wrap in try/except (#378) — companion to #377/v0.3.3. Fixes a positional-arg mismatch (call was missing tf_triggering_env_comparison, shifting every subsequent arg by one and routing to_date into the from_ts_ms slot → TypeError in compare.py:1300 → per-test for-loop aborted → tests stuck in tests_pending). Also wraps compute_regression_table in try/except so any future data-shape issue can't kill the stream.

Full Changelog: v0.3.3...v0.3.4

v0.3.3

Hotfix

• fix(coordinator): wrap per-test PR comment update in try/except (#377) — regression-ish: unwrapped update_comment_if_needed() in the per-test GitHub update path would propagate any transient GH API exception through the for-loop and abort the entire stream. Observed on PRs #14934 and #15018 on 2026-04-22 — streams got stuck at 1/N forever. Fixed by wrapping the is_actionable_pr block the same way the sibling prepare_regression_comment path was already wrapped.

Full Changelog: v0.3.2...v0.3.3