Log cksum prev by markhannum · Pull Request #6138 · bloomberg/comdb2

markhannum · 2026-08-14T00:44:24Z

.

roborivers

Cbuild submission: Success ✓.
Regression testing: Success ✓.

The first 10 failing tests are:
cdb2api_rollback_rc [failed with core dumped]
sc_resume_logicalsc_generated **quarantined**
cdb2jdbc
log_cksum_prev
diskspace_nollmeta
diskspace_nollmeta_nostripe_generated
consumer_non_atomic_default_consumer_generated **quarantined**
sc_downgrade [timeout] **quarantined**

roborivers

Cbuild submission: Success ✓.
Regression testing: Success ✓.

The first 10 failing tests are:
analyze_exit_immediately [failed with core dumped]
cdb2api_rollback_rc [failed with core dumped]
cdb2jdbc
diskspace_nollmeta_nostripe_generated
diskspace_nollmeta
consumer_non_atomic_default_consumer_generated **quarantined**
sc_downgrade [timeout] **quarantined**

roborivers

Cbuild submission: Success ✓.
Regression testing: Success ✓.

The first 10 failing tests are:
cdb2api_rollback_rc [failed with core dumped]
sc_resume_logicalsc_generated **quarantined**
cdb2jdbc
consumer_non_atomic_default_consumer_generated **quarantined**
manual_partition
sc_downgrade [timeout] **quarantined**
reco-ddlk-sql [timeout] **quarantined**
sql_logfill_autodisable [timeout]

roborivers

Cbuild submission: Success ✓.
Regression testing: Success ✓.

The first 10 failing tests are:
cdb2api_rollback_rc [failed with core dumped]
cdb2jdbc
consumer_non_atomic_default_consumer_generated **quarantined**
sc_downgrade [timeout] **quarantined**

roborivers

Cbuild submission: Success ✓.
Regression testing: Success ✓.

The first 10 failing tests are:
cdb2api_rollback_rc [failed with core dumped]
cdb2jdbc
sp_snapshot_generated
consumer_non_atomic_default_consumer_generated **quarantined**
sc_downgrade [timeout] **quarantined**

roborivers

Cbuild submission: Success ✓.
Regression testing: Success ✓.

The first 10 failing tests are:
sc_resume_logicalsc_generated **quarantined**
cdb2jdbc
consumer_non_atomic_default_consumer_generated **quarantined**
sc_downgrade [timeout] **quarantined**

roborivers

Cbuild submission: Success ✓.
Regression testing: Success ✓.

The first 10 failing tests are:
analyze_exit_immediately [failed with core dumped]
cdb2api_rollback_rc [failed with core dumped]
sc_resume_logicalsc_generated **quarantined**
cdb2jdbc
comdb2sys_pagesize_generated
consumer_non_atomic_default_consumer_generated **quarantined**
bind_query_plan
sc_downgrade [timeout] **quarantined**

roborivers

Cbuild submission: Success ✓.
Regression testing: Success ✓.

The first 10 failing tests are:
yast
consumer_non_atomic_default_consumer_generated **quarantined**
sc_downgrade [timeout] **quarantined**

markhannum · 2026-08-15T12:45:41Z

Performance: log_cksum_prev vs main

Benchmarked log-cksum-prev (09cbcfd) against main (7102927) on the same
machine, same build type, using a write-heavy workload aimed at the paths this
feature touches: the log-put critical section, per-record log volume, and
master→replicant shipping.

Result: no measurable throughput or latency regression. The feature's real
cost is log volume: +3.6% on a small-record schema (4 bytes × ~4 log records
per row). With log_cksum_prev off the log format is byte-identical to main.

Workloads

Table: { int a, int b } with one dup index on a.

W1 bulk: insert 1M rows in 100 transactions of 10k (server-side generate_series)
W2 concurrent commits: 8 parallel clients × 10k single-row autocommit
transactions (~16k commits/sec; stresses the serialized log-put path where
the previous-record checksum is stamped)
W3 updates: update 100k rows in 10 transactions of 10k
W4 log volume: du of the logs directory after W1 (flushed)

Single node (median of 3 runs, seconds; lower is better)

Workload	main	branch, cksum on	branch, cksum off
W1 bulk insert 1M rows	20.2	19.1	21.3
W2 8×10k concurrent commits	6.3	5.5	6.0
W3 update 100k rows	4.9	4.2	4.5
W4 log bytes after 1M rows	455.6 MB	471.8 MB (+3.6%)	455.6 MB

3-node cluster (mean of 2 runs, seconds)

Workload	main	branch, cksum on
W1 bulk insert 1M rows	64.4	58.9
W2 8×10k concurrent commits	16.1	15.0
W3 update 100k rows	6.2	6.3
W4 log bytes after 1M rows	455.7 MB	471.9 MB (+3.6%)

Notes

Run-to-run noise on this host was ±10–15% (worst single-node bulk rep: 31 s
on the branch, 23 s on main), so sub-10% throughput effects are below the
noise floor; the branch was as fast or faster than main in every workload's
median, so a regression hiding in the noise is unlikely. The +3.6% log
growth, by contrast, is exact and reproducible across every run: 16.2 MB per
1M rows = 4 bytes × ~4 log records per row. The percentage shrinks as
records get larger.
log_cksum_prev off produces log volume matching main to within a kilobyte,
i.e. the field is elided entirely, not zero-filled — legacy-defaults keep the
old format on disk.
Cluster numbers were taken with all three nodes co-hosted on one machine, so
they measure replication protocol overhead, not real network latency.

roborivers

Cbuild submission: Success ✓.
Regression testing: Success ✓.

The first 10 failing tests are:
consumer_non_atomic_default_consumer_generated **quarantined**
bind_query_plan
sc_downgrade [timeout] **quarantined**
reco-ddlk-sql [timeout] **quarantined**

roborivers

Cbuild submission: Success ✓.
Regression testing: Success ✓.

The first 10 failing tests are:
osql_cleanup [failed with core dumped]
consumer_non_atomic_default_consumer_generated **quarantined**
sc_downgrade [timeout] **quarantined**
truncatesc_offline_generated [timeout] **quarantined**
reco-ddlk-sql [timeout] **quarantined**

roborivers

Cbuild submission: Error ⚠.
Regression testing: Success ✓.

The first 10 failing tests are:
consumer_non_atomic_default_consumer_generated **quarantined**
sc_downgrade [timeout] **quarantined**

diskspace_nollmeta compared 'stat size' output verbatim, including the N% share each table holds of total space. That denominator counts the log files, so the ratio moves whenever the log format changes and the test fails for reasons that have nothing to do with table sizing. Strip the percentage and compare sizes only. Signed-off-by: Mark Hannum <[email protected]>

Rectype tags (ufid +1000, utxnid +2000) were decoded by open-coded range tests in ten files, each correct only for the tags that existed when it was written. Replace them with one decomposition. No behaviour change: normalize_rectype() agrees with the old version on every live rectype. Signed-off-by: Mark Hannum <[email protected]>

Will gate writing the previous log record's checksum into each record. Nothing reads it yet. READEARLY like utxnid_log, since a log-format tunable must be settled before the log opens. Off under legacy_defaults. Signed-off-by: Mark Hannum <[email protected]>

Adds rectype tag +4000: the record carries a u_int32_t prev_cksum after the utxnid in its common prefix. Reserved as zero here; the value is not known until the log region lock is held. In the prefix rather than at the end because generated readers get a recbuf with no length and can only parse forward. txn_auto.c tag tests move off "== DB___x + 2000", which the new tag breaks. Signed-off-by: Mark Hannum <[email protected]>

prev_cksum widens the common prefix, so hand-computed "4 + 4 + 8 + (utxnid_logged ? 8 : 0)" offsets are short by four bytes. Walking the prefix field by field was worse: it steps over only the fields it knows, so a new one is skipped silently -- __lc_cache_feed() lost prev_cksum that way. Consumers now read fixed fields at DB_REC_OFF_* and reach the body via __rectype_prefix_len(). Also fixes __log_put_next() reading __txn_dist_commit's generation at 4+4+8+4, which landed in the utxnid tail. Signed-off-by: Mark Hannum <[email protected]>

__log_put_next() stamps lp->last_cksum into the reserved field under the region lock; __log_putr() then advances it. Both dbt (our log) and udbt (what a master ships) are stamped. lp->last_cksum is in-memory only, re-derived by __log_recover() and __log_vtruncate() from the last real record at or before their position, crossing into the previous file when needed. LOGP headers carry log_size, which may differ between nodes, so they stay out of the chain. Also exposes prevcksum/cksum on comdb2_transaction_logs, and adds the record generator to llog_auto.c's CMake DEPENDS. Signed-off-by: Mark Hannum <[email protected]>

A chained record's checksum transitively covers the whole log prefix, so matching it byte-for-byte at the same LSN implies identical history -- the guarantee the commit-only rule provided. Decided from the record's own tag, so an unchained source falls back to commit-only. DB___txn_ckp_recovery stays excluded, and is never used for a generation: recovering a copy emits one above the cluster's generation. A non-commit anchor carries no generation, so find_match_lsn() asks the source for the one in effect there -- passing maxlsn as well as minlsn, or the vtab scans from the end of the log. A missing generation is only a bad logstream if the record should have carried one. Signed-off-by: Mark Hannum <[email protected]>

Checks each record's prevcksum against the previous record's cksum, in steady state, across a log file rollover, and across a bounce. pushlogs is asynchronous and master-only, so the rollover is polled for rather than assumed. Every node is flushed at the end: a flush that hangs is itself a bug, and failing here beats letting the harness's own silent flush time out and blame the datafiles. Signed-off-by: Mark Hannum <[email protected]>

roborivers

Cbuild submission: Error ⚠.
Regression testing: Success ✓.

The first 10 failing tests are:
ssl_san
consumer_non_atomic_default_consumer_generated **quarantined**
ssl_set_cmd
ssl_prefer
ssl_dbname
sc_downgrade [timeout] **quarantined**
reco-ddlk-sql [timeout] **quarantined**

markhannum added the WIP Work in Progress label Aug 14, 2026

roborivers approved these changes Aug 14, 2026

View reviewed changes

markhannum force-pushed the log-cksum-prev branch 8 times, most recently from 0a1e39c to 5b57ebf Compare August 14, 2026 10:33

roborivers approved these changes Aug 14, 2026

View reviewed changes

markhannum force-pushed the log-cksum-prev branch 2 times, most recently from 46cfbfd to 7f10ca4 Compare August 14, 2026 11:56

roborivers approved these changes Aug 14, 2026

View reviewed changes

markhannum force-pushed the log-cksum-prev branch from 7f10ca4 to 09cbcfd Compare August 14, 2026 15:42

roborivers approved these changes Aug 14, 2026

View reviewed changes

markhannum force-pushed the log-cksum-prev branch from 97a2290 to b7aa972 Compare August 15, 2026 12:43

roborivers approved these changes Aug 15, 2026

View reviewed changes

markhannum removed the WIP Work in Progress label Aug 16, 2026

markhannum force-pushed the log-cksum-prev branch from b7aa972 to 573f1db Compare August 17, 2026 17:55

roborivers approved these changes Aug 18, 2026

View reviewed changes

markhannum force-pushed the log-cksum-prev branch from 573f1db to 0f62fbf Compare August 18, 2026 20:19

roborivers suggested changes Aug 19, 2026

View reviewed changes

markhannum added 2 commits August 21, 2026 12:31

markhannum added 6 commits August 21, 2026 12:31

markhannum force-pushed the log-cksum-prev branch from 0f62fbf to 2297c47 Compare August 21, 2026 16:33

roborivers suggested changes Aug 21, 2026

View reviewed changes

Navigation Menu

Uh oh!

Log cksum prev - #6138

Log cksum prev#6138
markhannum wants to merge 8 commits into
bloomberg:mainfrom
markhannum:log-cksum-prev

markhannum commented Aug 14, 2026

Uh oh!

roborivers left a comment

Uh oh!

roborivers left a comment

Uh oh!

roborivers left a comment

Uh oh!

roborivers left a comment

Uh oh!

roborivers left a comment

Uh oh!

roborivers left a comment

Uh oh!

roborivers left a comment

Uh oh!

roborivers left a comment

Uh oh!

markhannum commented Aug 15, 2026

Uh oh!

roborivers left a comment

Uh oh!

roborivers left a comment

Uh oh!

roborivers left a comment

Uh oh!

roborivers left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

markhannum commented Aug 14, 2026

Uh oh!

roborivers left a comment

Choose a reason for hiding this comment

Uh oh!

roborivers left a comment

Choose a reason for hiding this comment

Uh oh!

roborivers left a comment

Choose a reason for hiding this comment

Uh oh!

roborivers left a comment

Choose a reason for hiding this comment

Uh oh!

roborivers left a comment

Choose a reason for hiding this comment

Uh oh!

roborivers left a comment

Choose a reason for hiding this comment

Uh oh!

roborivers left a comment

Choose a reason for hiding this comment

Uh oh!

roborivers left a comment

Choose a reason for hiding this comment

Uh oh!

markhannum commented Aug 15, 2026

Performance: log_cksum_prev vs main

Workloads

Single node (median of 3 runs, seconds; lower is better)

3-node cluster (mean of 2 runs, seconds)

Notes

Uh oh!

roborivers left a comment

Choose a reason for hiding this comment

Uh oh!

roborivers left a comment

Choose a reason for hiding this comment

Uh oh!

roborivers left a comment

Choose a reason for hiding this comment

Uh oh!

roborivers left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants