Hi, Prometheus[0] author here. Thanks for the interesting article!
Since I was curious how this compares to Prometheus's internal storage for writes, I whipped up some (disclaimer: very naive and ad-hoc!) benchmarks[1] to get a rough feeling for Catena's performance. I am not achieving a lot of write performance with it yet, but maybe I'm doing something wrong or using it inefficiently. Some questions to investigate would be: what's the best number of rows to batch in one insert, and are timestamps in seconds, milliseconds, or essentially only user-interpreted (I noticed the partitioning at least depends heavily on the interval between timestamps)? So far I've just done a tiny bit of fiddling and results haven't changed dramatically.
The benchmark parameters:
* writing 10000 samples x 10000 metrics (100 million data points)
* initial state: empty storage
* source names: constant "testsource" for all time series
* metric names: "testmetric_<i>" (0 <= i < 10000)
* values: the metric index <i> (constant integer value within each series)
* timestamps: starting at 0 and increasing by 15 seconds for every iteration
* Other machine load: SoundCloud playing music in the background
The benchmark results:
#### Prometheus ####
(GOMAXPROCS=4 go run prometheus_bench.go -num-metrics=10000 -samples-per-metric=10000)
Time: 1m26s
Space: 138MB
#### Catena ####
(GOMAXPROCS=4 go run catena_bench.go -num-metrics=10000 -samples-per-metric=10000)
Time: 1h25m
Space: 190MB
So in this particular benchmark Catena took 60x longer and used 1.4x more space.
Please don't take this as discouragement or a statement on one being better than the other. Obviously Catena is very new and also probably optimized for slightly different use cases. And possibly I'm just doing something wrong (please tell me!). I also haven't dug into possible performance bottlenecks yet, but I saw it utilize 100% of all 4 CPU cores the entire time. In any case, I'd be interested in a set of benchmarks optimized specifically for Catena's use case.
Unfortunately we also haven't fully documented the internals of Prometheus's storage yet, but a bit of background information can be found here: http://prometheus.io/docs/operating/storage/ Maybe that's worth a blog post sometime.
Thanks for trying it out! I haven't had time to run any benchmarks, so I really appreciate you taking the time to do this (especially since it took so long!).
I'm not sure what the best batch size is at the moment. Timestamps are int64s, and it's up to the user to interpret them as they wish. Partition sizes are in terms of the number of timestamps. If you had timestamps which correspond to seconds, and you wanted each partition to be 1 day, you'd choose 86400. This isn't configurable yet unless you modify the source.
I'm not surprised it's that slow. I'm not a storage engine expert, still a college student, and this has my first lock-free list implementation :). There is a lot of silliness going on with inserts, like launching a goroutine for each metric on inserts[0], using lock-free O(n) lists for sources and metrics[1] when I could have used a map, there's still a lock left that should be removed[2], and a bunch of others.
On an unrelated note, I see that someone from Prometheus will be at Monitorama. I'll be there too, so I'd love to you guys some more.
Ah ok, so some obvious things left to optimize then :) Thanks for the explanations!
As far as I know, none of the Prometheus core developers will be at Monitorama this year. But we'll be at SRECon Dublin, DevOpsCon Berlin, GopherCon Denver, and hopefully also at one (or both) of the VelocityConfs later this year.
Indeed, it would be great to catch up in person if you're at any of these.
Since I was curious how this compares to Prometheus's internal storage for writes, I whipped up some (disclaimer: very naive and ad-hoc!) benchmarks[1] to get a rough feeling for Catena's performance. I am not achieving a lot of write performance with it yet, but maybe I'm doing something wrong or using it inefficiently. Some questions to investigate would be: what's the best number of rows to batch in one insert, and are timestamps in seconds, milliseconds, or essentially only user-interpreted (I noticed the partitioning at least depends heavily on the interval between timestamps)? So far I've just done a tiny bit of fiddling and results haven't changed dramatically.
The benchmark parameters:
* writing 10000 samples x 10000 metrics (100 million data points)
* initial state: empty storage
* source names: constant "testsource" for all time series
* metric names: "testmetric_<i>" (0 <= i < 10000)
* values: the metric index <i> (constant integer value within each series)
* timestamps: starting at 0 and increasing by 15 seconds for every iteration
* GOMAXPROCS=4 (4-core "Core i5-4690K" machine, 3.5GHz)
* Disk: SSD
* Other machine load: SoundCloud playing music in the background
The benchmark results:
#### Prometheus ####
(GOMAXPROCS=4 go run prometheus_bench.go -num-metrics=10000 -samples-per-metric=10000)
Time: 1m26s Space: 138MB
#### Catena ####
(GOMAXPROCS=4 go run catena_bench.go -num-metrics=10000 -samples-per-metric=10000)
Time: 1h25m Space: 190MB
So in this particular benchmark Catena took 60x longer and used 1.4x more space.
Please don't take this as discouragement or a statement on one being better than the other. Obviously Catena is very new and also probably optimized for slightly different use cases. And possibly I'm just doing something wrong (please tell me!). I also haven't dug into possible performance bottlenecks yet, but I saw it utilize 100% of all 4 CPU cores the entire time. In any case, I'd be interested in a set of benchmarks optimized specifically for Catena's use case.
Unfortunately we also haven't fully documented the internals of Prometheus's storage yet, but a bit of background information can be found here: http://prometheus.io/docs/operating/storage/ Maybe that's worth a blog post sometime.
[0] http://prometheus.io/
[1] The code for the benchmarks is here: https://gist.github.com/juliusv/ce7c3b5368cd7adf8bc6