Not a throughput tuning knob — it is a blast-radius limit. A queue message on this topic lives 24
hours (deriveRetentionSeconds caps retention at the dedup window, which is itself capped at 24h),
so any burst larger than the topic can drain in a day loses its tail SILENTLY: the message just
expires, leaving no failed row and no failed job_executions record to notice. At the topic's
maxConcurrency of 50 this ceiling drains in minutes, so it is never the binding constraint for a
real patient — it exists so a bulk caller cannot accidentally push a whole corpus through one call.
Bulk callers are expected to page against this rather than raise it.
The most charts one call may enqueue.
Not a throughput tuning knob — it is a blast-radius limit. A queue message on this topic lives 24 hours (
deriveRetentionSecondscaps retention at the dedup window, which is itself capped at 24h), so any burst larger than the topic can drain in a day loses its tail SILENTLY: the message just expires, leaving no failed row and no failedjob_executionsrecord to notice. At the topic'smaxConcurrencyof 50 this ceiling drains in minutes, so it is never the binding constraint for a real patient — it exists so a bulk caller cannot accidentally push a whole corpus through one call.Bulk callers are expected to page against this rather than raise it.