Firebird: Developer's Report: QA Stress Testing & Tools Development

February - April 2014

February 2014

Benchmarked inserts in v3.0 vs v2.5: compare cases when adding 1 mil rows using 10 attaches at the same time

Result: v3.0 is significantly faster generally, but GEN_ID is much slower in v3.0

Tested heavy index load along with nbackup in v3.0 vs v2.5

Result: MON$ tables are inaccessible untill nbackup -N is finished

Cancelling of gfix -sweep via delete from mon$attachments seems not working: gfix process is still alive

Result: CORE-4337

Non-privileged user can: (1) grant to himself other's role and (2) update rdb$security_classes table

Result: CORE-4341, CORE-4342

Benchmarked of bulk inserts into indexed table, single connect, 9999 execute blocks (1000 rows each of them)

Result: no noticable differences between v3.0 and v2.5

Benchmarked simultaneous restore of a production database using services

Result: no noticable differences between v3.0 and v2.5

Benchmarked hash join in 3.0 vs merge join in 2.5; also hash join vs nested loop join (both in 3.0)

Result: hash join in almost all cases runs with the same or better performance

Found issue when restore of a production database could not be finished on v2.5 but succeeds in v3.0

Result: CORE-4352

Compare performance between nested loops in FB 3.0 vs 2.5 in simple query that counts number of 'lucky tickets'

Result: v3.0 looses about 15% in performance

Resume checking accuracy with hash join in plan

Result: no errors in result sets, hash join is same or faster than merge join

Finish comparing performance and accuracy of complex queries in FB v3.0 vs MS SQL 2005

Result: no errors found in result sets, FB is a worse performer in medium and big data volumes

March 2014

Found a few cases when cursor stability does not work as expected

Result: added test cases to CORE-3362

Asynchronous cancellation request can be ignored in some cases

Result: CORE-4356

Non-priviledged user can block creation of new tables by select ... from rdb$database FOR UPDATE WITH LOCK

Result: CORE-4357

GROUP BY from derived table which contains group by on field with LITERAL value returns wrong result

Result: CORE-4360, fixed by Dmitry Yemanov

Investigated stored functions: found some issues about no-caching in case of DETERMINISTIC clause

Result: reported to core developers

Poor performance of windowed lead()over() function in comparison with explicit cursor (not uses indexes)

Result: reported to core developers

Tested index scans vs full table scans

Result: index scan wins when it covers up to 97% of data

Wrong result of WHERE predicate when it contains NULL IS NOT DISTINCT FROM (select min(NULL) from ...)

Result: CORE-4366, fixed by Dmitry Yemanov

Poor performance of explicit cursors containing correlated subqueries in the select list

Result: CORE-4379, fixed by Dmitry Yemanov

Start implementing big test that emulates OLTP activity

April 2014

Continued implementing OLTP-emulation test, refactored it multiple times in order to increase performance and minimize lock conflicts frequency

SELECT WITH LOCK may enter an infinite loop for a single record

Result: CORE-4388, fixed by Dmitry Yemanov

Trigger on DISCONNECT does not fire in FB 3.0

Result: regression is fixed by Dmitry Yemanov

Inner join (nested loop) does not take in account predicate in WHERE clause when driven source is VIEW based on GROUP BY

Result: CORE-4399

MERGE set row_count always to 1 (despite of actual rows affected)

Result: CORE-4400

Restore troubles: ERROR:Access to UDF/BLOB-filter module "" is denied by server administrator

Result: supposedy fixed by Dmitry Yemanov

FB crashes when fbtracemgr restarts

Result: fixed by Alex Peshkov

Pavel Zotov
Moscow, Russian Federation

Join Firebird!

Follow Us

Newsletter