Gabriel39 commented on issue #70251:
URL: https://github.com/apache/doris/issues/70251#issuecomment-7001561239

   D10 performance evidence is available for candidate ce3c52dc4fe; this does 
not complete the D10 soak/fault gates.
   
   - Release BE (-O3, SANITIZE=OFF), matching native FE plugin, 1 FE / 1 BE; 
CPU affinities cover 16 BE cores, 8 FE cores and 16 source cores on a shared 
host. MongoDB is co-located, authenticated and warm-cache. Main fixtures use 
MongoDB 7.0.28/8.0.20 replica sets and two replicated shards through mongos.
   - Deterministic 250,000/1,000,000-row collections; full BSON row size646 
bytes. Each of22 scenarios has one warmup and seven measured repetitions per 
environment, in randomized blocks. No slow measured samples are removed. Query 
caches and extra profiling are disabled during latency timing; source 
diagnostics and six-repeat FE planning measurements are separate.
   
   | Source | Projected single split, seconds | Range16, seconds | Indexed 
predicate speedup over projection-only |
   | --- | ---: | ---: | ---: |
   | 7.0.28 replica set | 1.516 | 2.533 | 2.93x |
   | 7.0.28 mongos | 2.017 | 2.040 | 3.89x |
   | 8.0.20 replica set | 1.523 | 3.044 | 2.92x |
   | 8.0.20 mongos | 2.518 | 3.052 | 5.78x |
   
   The ordinary-collection narrow aggregate shows no general range-scan benefit 
here: the largest range contains996,178 of1,000,000 rows, and source documents 
examined rise from1m to2m because of skew and complement work. Wide-field 
performance varies by topology and is reported separately. Projection reduces 
effective BSON bytes by15.38x. The predicate comparison uses NULL mismatch 
policy on valid data; the default ERROR policy is retained as a separate 
control and intentionally keeps filtering local.
   
   Additional checks:9 DATETIME(0/3/6) x Decimal precision/scale combinations, 
basic reads of ordinary collections/computed no-id views/logical time-series, 
and time-series Date filtering. All817 formal completed queries match 
independent expected outputs;103 warmups and142 separate diagnostic/planning 
queries are excluded from formal latency statistics. Explicit Date/Decimal 
precision-loss guards verify rejection or NULL instead of silent rounding. The 
precision performance data uses identical whole-second Dates under ERROR policy 
and requires both non-null column counts to equal1m.
   
   A same-library libmongoc2.5.8 microexperiment measures about495ms per pool 
destruction versus5–6ms to read1,000 indexed rows. This explains an important 
current lifecycle cost; the previously agreed D06 scope is unchanged. Final 
cleanup checks cover24 source endpoints with zero test-data cursors, including 
a positive control for the cursor detector.
   
   One auxiliary fixture preparation exhausted its4GiB WiredTiger history-store 
volume after the main matrix had finished. The affected auxiliary work was 
rerun sequentially after rebuilding that replica set with16GiB tmpfs and 
grouping time-series loading by metadata; this event and the corrected 
precision preflight are retained in the report. No product code was changed.
   
   A PDF report, raw numeric CSV/JSON and sanitized reproduction scripts are 
delivered separately. Shared-host/co-located warm-cache numbers are relative 
baselines, not production throughput or P95/P99 promises. Cross-machine/TLS 
performance, 1,000 mixed query/cancel/failure cycles and the24-hour 
mixed-workload soak remain unverified; keep D10 open.
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to