Open dmarriner opened 1 year ago
Hey @dmarriner!
I'm peaceful with this. The probability a map
iteration would pass all 100 asserts is very low if I'm not mistaken. Assuming that iteration is truly random:
(1 0.5 0.5) ^ (100) = 6.22301528e-61
One nit: could you update that PR comment to have
Summary Sort the labels on a given metric before adding them to the cache.
Motivation This fixes a bug in veneur-prometheus that caused diff calculations to sporadically break when multiple labels were added via the -a flag. Labels were stored as key/values in a map and returned in a potentially different order each time since map iteration order is random in golang. This caused the translator to sometimes emit the cumulative value instead of the diffed value, because it would seem as though the metric was new when in fact it had been previously cached with a different label ordering. For example, the translator might look for "counter-key2:value2-key1:value1" when it had been previously cached as "counter-key1:value1-key2:value2"
It'll make viewing git blame history a bit easier.
Updated the PR comment!
Summary
Sort the labels on a given metric before adding them to the cache.
Motivation
This fixes a bug in veneur-prometheus that caused diff calculations to sporadically break when multiple labels were added via the -a flag. Labels were stored as key/values in a map and returned in a potentially different order each time since map iteration order is random in golang. This caused the translator to sometimes emit the cumulative value instead of the diffed value, because it would seem as though the metric was new when in fact it had been previously cached with a different label ordering. For example, the translator might look for "counter-key2:value2-key1:value1" when it had been previously cached as "counter-key1:value1-key2:value2"
Test plan
Wrote an integration test.
Rollout/monitoring/revert plan
N/A