Gabriel39 commented on issue #70251:
URL: https://github.com/apache/doris/issues/70251#issuecomment-7007130197

   Performance follow-up for [PR 
#70708](https://github.com/apache/doris/pull/70708), commit 
`9905e03498da1f46bde9a35906ad39ee20a879ce`.
   
   FE now uses bounded source `$sample` at the common T. Complete BSON-order 
`_id` intervals remove the independent complement pass; custom missing/multikey 
fields retain their typed complement. Single remains the default, existing scan 
controls are reused, and D06/shared StopFlag changes remain excluded.
   
   Fresh paired measurements use the same immutable 250k/1m inputs, one FE/one 
Release BE and MongoDB 7.0.28/8.0.20 replica-set/mongos fixtures. Across both 
phases: 1,232 measured single queries, 176 warmups, 176 separate source 
diagnostics, 96 planning profiles and 192 four-client queries. Every result 
passes independent arithmetic/Decimal checks; slow samples are retained. For 1m 
rows, examined data documents drop from 2m to 1m and largest split share drops 
from 50.03–99.62% to 6.83–7.23%.
   
   | MongoDB | range16 before (s) | range16 after (s) | single before → after 
(s) | largest split before → after |
   |---|---:|---:|---:|---:|
   | 7 replica set | 3.043 | 0.541 | 2.016→1.514 | 99.62% → 6.83% |
   | 7 mongos | 4.568 | 0.570 | 2.521→2.517 | 99.62% → 7.23% |
   | 8 replica set | 3.045 | 2.599 | 1.526→2.518 | 99.62% → 6.90% |
   | 8 mongos | 3.551 | 2.661 | 3.021→3.517 | 50.03% → 6.98% |
   
   
   Observed costs: MongoDB8 indexed-range medians rise (8 replica set 
0.547→0.619s; 8 mongos 0.573→0.674s). Four-client MongoDB8 mongos range16 QPS 
changes 0.625→0.519 (two batches per phase). Single controls also drift; these 
observations are retained without causal attribution or a saturation-throughput 
claim.
   
   The new report includes all 22 workloads, medians/IQRs, single controls, 
CPU/source/planning counters, RSS observations and numerical data with a 
standard-library recomputation script. Shared-host phase drift prevents 
treating every timing difference as a causal gain; MongoDB 8 replica-set 
range16 remains close to single.
   
   Local final verification: 148 Mongo FE UTs, full FE Checkstyle/build, four 
retained split SQL runs, 1,568 source-backed multisets, 173,088 independent 
precision cells, 12 transactional exact-T comparisons, 510 controlled 
read-parameter records and 24 real two-BE checks pass. DATETIME(0..6), nine 
Decimal precision/scale pairs, four timezones and extreme/mixed BSON `_id` 
values are covered. Six new framework-generated golden lines add two 601/180300 
tags; all 16 previous tags are unchanged.
   
   Qualification limitation: MongoDB 7 default nearest hedging left a losing 
shard cursor with no mongos owner. Eighty fully consumed source-only PyMongo 
reads independently reproduce this behavior. The parameter cleanup matrix 
temporarily disables/restores router hedging and retains strict zero-cursor 
checks; default hedged source cursor cleanup remains unqualified. Product 
routing defaults are unchanged. Remote CI/review is pending.
   
   D10 remains open: this report does not establish cross-machine/TLS 
performance, saturation throughput, stable memory, a 24-hour soak or 
1,000-cycle mixed-fault qualification. The full PDF/HTML and sanitized data 
bundle were delivered separately.
   


-- 
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