← All use cases
The scenario
The backfill is running. Twenty seconds in you realise it is pointed at the
staging table, or the date filter was wrong, or it is one job per row instead of
one job per batch.
Every queued chunk is an LLM call you are paying for, and it will keep going
until it finishes.
What this recipe shows that no other does: operational control of work
already in flight — cancelling it, and tearing down the plumbing around it.
Step 1 — Cancel the job
The two numbers matter and they are not the same thing. Cancelled parts stop.
Running parts finish — work already handed to a worker is not interrupted, so
a large job keeps producing memories for a little while after you cancel it.
Watch, do not assume.
Step 2 — Find out what landed
Step 3 — Undo the damage
If the import was file-based, deleting the document takes its memories with it:
If it was a write of loose records and the space existed only for the import,
the honest cleanup is to
delete the space and start again. Picking
individual bad memories out of a good space is slow and error-prone — which is
an argument for importing into a scratch space first.
Step 4 — Turn off what was watching it
A cancelled import usually leaves a webhook that no longer has a job to report on.
Deleting stops queued deliveries for it — worth doing before your endpoint gets
a burst of events about a job you abandoned.
Evals
Practise this before you need it, on a space you do not care about:
- Queue a large batch, cancel it after a few seconds.
- Record
cancelled_parts and running_parts.
- Poll
list_memories until total stops moving. The gap between cancelling
and stopping is the number worth knowing — it is what you will be asked.
- Confirm cleanup actually emptied it.
Guardrails
Cancelling does not refund what already ran. Extraction that has happened
is billed. The saving is on the part that had not started, which is why the
first twenty seconds matter more than the next twenty minutes.
- Import into a scratch space. It converts “pick out the bad memories” into
“delete the space”, and that is the whole difference between a bad afternoon
and a bad minute.
- A cancelled job is a terminal state, not a pause. There is no resume;
re-import what you still want.
- Throughput is bounded platform-side, so a large import runs for a long
time. That is exactly why the kill switch is worth knowing before you need it.
