MariaDB 11.8.10, MyRocks. 2026-11-21, Europe/Paris.
Five distinct fingerprints returned together by the digest example loader; observations A-E. Each variant ran 50 times for each example, in shuffled balanced rounds. Random order is reproducible (MT19937). First runs, slow runs and regressions are included. Query result cache disabled. No application deployment.
Production reads (first series): seed 20261121; PHP mysqli, one connection, consistent read-only snapshots, SHOW PROFILES in milliseconds. Replay tests a real existing point. Append uses its epoch plus 86400 and a date one day later. Second SQL-only series: seed 21112026, separate control.
Full application transactions: seed 202611215; elapsed time from PHP hrtime around the actual operation, including named lock, COMMIT and release. Before and after run in independent databases with the same historical keys for the five fingerprints (28,463 points at snapshot), 45 partitions, synthetic payloads, 4 healthy streams and 1 in alarm. Append creates one new point per round per stream. Final state compared. Cold fixtures start with no retained history, either new or with an existing healthy legacy watermark. Each cold round resets the fixture outside the timer. The first application comparison runs V1 and V2. An additional three-way preflight before switching the PHP fingerprint literal to the standard escaping helper produced identical SQL and is preserved as preflight-*.csv. The exact final code was rerun. The final campaign in transactions3.csv and cold3.csv compares V1 (before), V2 (middle) and V3 (after) in randomized order against identical copies, checking the decisions and end state of all three. The cost of restoring historical alert occurrences was not measured.
Physical SQL layouts were compared on identical MariaDB/MyRocks copies. Seed 219117. First comparison: 571957 keys (all 543719 rows from November 20 and the five complete histories). Extended comparison: 621865 keys (49908 additional background rows from November 21; first 20000 extracted rows, all rows from 22:00 to the saved example time 22:29:19, and every 32nd remaining row). Payloads excluded. Primary key and two existing secondary indexes retained in every layout. ORDER BY output equality checked on every round. A time index alone can respond differently on sparser fingerprints than in this active sample. Four full layout row counts were verified. Timings from distinct campaigns or from laboratory and production do not have a shared baseline.
The isolated transaction measurements and live read-path measurements time different operations in different environments. Profiling overhead, buffer state, server load and five recent streams limit extrapolation.

exemples.csv maps the five published observations to their detector codes, states and retained point counts. plans-parcours.csv contains supplemental EXPLAIN PARTITIONS outputs from 2026-11-21 after the timing runs. analyze-time-zone.json reports iterator counters only, from one read-only execution of each query.
