mirror of
https://github.com/timescale/timescaledb.git
synced 2025-05-19 20:24:46 +08:00
ac97c564e2
When two CAggs on the same hypertable are refreshed at the same type, we had a race condition on the invalidation threshold table occur. So far, the table has been locked with a non-self-conflicting lock. Therefore, both scanners ran at the same time, but only one was able to lock the threshold value with a proper tuple lock. The other scanner ignored this failing lock and just returned. Therefore, the field computed_invalidation_threshold was never populated and still contains 0. So, invalidation_threshold_set_or_get returns and refresh end interval of 0. As a consequence, the `if (refresh_window.start >= refresh_window.end)` branch in continuous_agg_refresh_internal could be taken and we return from the refresh without doing any work. This patch adds proper error reporting and also implements some retry logic to avoid these problems. A self-conficting lock is not used due to the problems discussed in #5809.