I've been looking into possible alternatives, based on the ideas given here.
I) First one completely different approach might be: - alert: target-down expr: 'max_over_time( up[1m0s] ) == 0' for: 0s and: ( - alert: single-scrape-failure expr: 'min_over_time( up[2m0s] ) == 0' for: 1m or - alert: single-scrape-failure expr: 'resets( up[2m0s] ) > 0' for: 1m or perhaps even - alert: single-scrape-failure expr: 'changes( up[2m0s] ) >= 2' for: 1m (which would however behave a bit different, I guess) ) plus an inhibit rule, that silences single-scrape-failure when target-down fires. The for: 1m is needed, so that target-down has a chance to fire (and inhibit) before single-scrape-failure does. I'm not really sure, whether that works in all cases, though, especially since I look back much more (and the additional time span further back may undesirably trigger again. Using for: > 0 seems generally a bit fragile for my use-case (because I want to capture even single scrape failures, but with for: > 0 I need t to have at least two evaluations to actually trigger, so my evaluation period must be small enough so that it's done >= 2 during the scrape interval. Also, I guess the scrape intervals and the evaluation intervals are not synced, so when with for: 0s, when I look back e.g. [1m] and assume a certain number of samples in that range, it may be that there are actually more or less. If I forget about the above approach with inhibiting, then I need to consider cases like: ----time----> - 0 1 0 0 0 0 0 0 first zero should be a single-scrape-failure, the last 6 however a target-down - 1 0 0 0 0 0 1 0 0 0 0 0 0 same here, the first 5 should be a single-scrape-failure, the last 6 however a target-down - 1 0 0 0 0 0 0 1 0 0 0 0 0 0 here however, both should be target-down - 1 0 0 0 0 0 0 1 0 1 0 0 0 0 0 0 or 1 0 0 0 0 0 0 1 0 0 1 0 0 0 0 0 0 here, 2x target-down, 1x single-scrape-failure II) Using the original {min,max}_over_time approach: - min_over_time(up[1m]) == 0 tells me, there was at least one missing scrape in the last 1m. but that alone would already be the case for the first zero: . . . . . 0 so: - for: 1m was added (and the [1m] was enlarged) but this would still fire with 0 0 0 0 0 0 0 which should however be a target-down so: - unless max_over_time(up[1m]) == 0 was added to silence it then but that would still fail in e.g. the case when a previous target-down runs out: 0 0 0 0 0 0 -> target down the next is a 1 0 0 0 0 0 0 1 -> single-scrape-failure and some similar cases, Plus the usage of for: >0s is - in my special case - IMO fragile. III) So in my previous mail I came up with the idea of using: - alert: target-down expr: 'max_over_time( up[1m0s] ) == 0' for: 0s - alert: single-scrape-failure expr: 'min_over_time(up[15s] offset 1m) == 0 unless max_over_time(up[1m0s]) == 0 unless max_over_time(up[1m0s] offset 1m10s) == 0 unless max_over_time(up[1m0s] offset 1m) == 0 unless max_over_time(up[1m0s] offset 50s) == 0 unless max_over_time(up[1m0s] offset 40s) == 0 unless max_over_time(up[1m0s] offset 30s) == 0 unless max_over_time(up[1m0s] offset 20s) == 0 unless max_over_time(up[1m0s] offset 10s) == 0' for: 0m The idea was, that when I don't use for: >0s, the first time window where one can be really sure (in all cases), that whether it's a single-scrape-failure or target-down is a 0 in -70s to -60s: -130s -120s -110s -100s -90s -80s -70s -60s -50s -40s -30s -20s -10s 0s/now | | | | | | | 0 | | | | | | | | | | | | | | | | | | 1 | 0 | 1 | case 1 | | | | | | | 0 | 0 | 0 | 0 | 0 | 0 | 0 | case 2 | | | | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | case 3 In case 1 it would be already clear when the zeros is between -20 and -10. But if there's a sequence of zeros, it takes up to -70s to -60s, when it becomes clear. Now the zero in that time span could also be that of a target-down sequence of zeros like in case 3. For these cases, I had the shifted silencers that each looked over 1m. Looked good at first, though there were some open questions. At least one main problem, namely it would fail in e.g. that case: -130s -120s -110s -100s -90s -80s -70s -60s -50s -40s -30s -20s -10s 0s/now | 1 | 1 | 1 | 1 | 1 | 1 | 0 1 | 0 | 0 | 0 | 0 | 0 | 0 | case 8a The zero between -70s to 60s would be noticed, but still be silenced, because the one would not. Chris Siebenmann suggested to use resets(). ... and keep_firing_for:, which Ben Kochie, suggested, too. First I didn't quite understand how the latter would help me? Maybe I have the wrong mindset for it, so could you guys please explain what your idea was wiht keep_firing_for:? IV) resets() sounded promising at first, but while I tried quite some variations, I wasn't able to get anything working. First, something like resets(up[1m]) >= 1 alone (with or without a for: >0s) would already fire in case of: ----time----> 1 0 which still could become a target-down but also in case of: 1 0 0 0 0 0 0 which is a target down. And I think even if I add some "unless ..." I'd still have the problem as above in (II), that I get a false positive alert, when a true target-down sequence moves through. So just like in (III) I'd need do that shifted silencers. resets(up[1m]) >= 2 wouldn't work either e.g. in case of: 1 0 1 1 1 1 1 1 there simply is no 2nd reset. I even tried a variant where the target-down must come first in the rules definition: - alert: target-down expr: 'up == 0' for: 1m <- for is needed here, or I get no ALERTS - alert: single-scrape-failure expr: 'resets(up[1m0s]) > 0 unless on (instance,job) ALERTS{alertname="target-down"}' for: 0m and where I then used ALERTS trying to filter ... but no success. V) Instead of resets() I tried changes() (which is even not only defined for counters): - alert: target-down expr: 'max_over_time( up[1m0s] ) == 0' for: 0s - alert: single-scrape-failure expr: using just changes(up[1]) >= 1 does of course not work, as it could be an incoming target-down 1 0 0 0 0 0 0 or an outgoing one: 0 0 0 0 0 0 1 using changes(up[1]) >= 2 seems promising first, if I have e.g. 1 1 1 1 0 1 it's already clear, that it's a single-scrape-failure... but it could be something like 0 0 0 0 0 0 1 1 0 0 0 i.e. an outgoing target-down and something that may still become one. using changes(up[1m5s]) >= 2 unless max_over_time(up[1m0s] offset 1m) == 0 unless max_over_time(up[1m0s] offset 50s) == 0 unless max_over_time(up[1m0s] offset 40s) == 0 unless max_over_time(up[1m0s] offset 30s) == 0 unless max_over_time(up[1m0s] offset 20s) == 0 used the above, and filtered again the shifted 1m time spans (no need to look at offset 0s or 10s). But that fails e.g. in the case of 0 0 0 0 0 0 1 0 1 1 1 1 1 1 1 (i.e. a target-down followed by a single-scrape-failure followed by OK) VI) avg_over_time. I guess I might just not understand what you mean, but at least something like: expr: 'avg_over_time(up[1m10s]) < 1 and avg_over_time(up[1m10s]) > 0' for: 1m fails already in the simple case of 0 0 0 0 0 1 where it gives a false alert after the target-down Well... guess I'm a my wits' end and this might simply not be possible with PromQL. Cheers, Chris. -- You received this message because you are subscribed to the Google Groups "Prometheus Users" group. To unsubscribe from this group and stop receiving emails from it, send an email to prometheus-users+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/prometheus-users/1bc1ade0-b28e-440d-ab0c-690cec0010edn%40googlegroups.com.