QuestDB
QuestDB, a database, power cut at forty different points
Summary
Of 80 power cuts on a real QuestDB, 5 ended in a silent loss: saved writes made after the restart were counted as lost and nothing marked the table. Molina made the cuts at 40 points of one workload and reproduced a loss. A public QuestDB issue already describes that loss, and the count is from one sweep, so it is not a rate.
The full experimental record
What I ran
The plan was 80 power cuts on a real QuestDB 10.0.1: 40 points of the workload, each cut on a disk that keeps every write that reached it and on a disk with a volatile cache. In 78 of them the power went at the request the plan asked for. The output does not say which cuts did not fire, and they are counted in the total. 5 were what the gate calls a silent loss: the table was not suspended, and the commits made after the restart are counted as lost.
What it shows
MIGHT. In 6 of 6 runs of QuestDB restarted on the crash image of one cut that left the writer ahead of the sequencer, the gate counted every commit made after the restart as lost, with the table not marked suspended.
What it does not show
That this bug is new. The gate calls the loss the one that a QuestDB issue describes. According to the repository's notes (docs/questdb-in-a-world.md), that issue was open and public when the notes were written. What this result adds is a reproduction from power cuts the hypervisor made.
A rate. This result reports one sweep and does not say how often a disk loses commits. According to the repository's notes (docs/questdb-in-a-world.md), earlier sweeps of this gate gave other counts on each kind of disk, which this result does not carry.
That the rate here is the rate in production.
That the sweep is exhaustive, that other versions behave the same, or that a finding here is a finding about QuestDB in production.
Whether the other silent cases, and the restarts of the kept image, also had a read-back that failed. The output does not say whether the other silent cases or the restarts did.
Conditions
One version (10.0.1), one image, pulled by digest. The gate shortened the guest kernel's writeback timers so that a short run has disk traffic.
Where it comes from
evidence/20261104T225051Z-kvm-questdb.out
This page is the short form of the record. The finding and its limits come first, and I left out what did not fit.