Production postmortemThe case of 99.99% percentile
A customer called us with a problem, for the most part, RavenDB was very well behaved, but they were worried about the 99.99% percentile of request duration. While the 99.9% was excellent (around a few milliseconds), the 99.99% was measured in seconds.
Usually, this is a particularly slow request that happens, and we can handle that by figuring out what is slow on that particular request and fix it. Common causes for such an issue is a request that returns a lot of unnecessary data, such as large full documents, when it needs just a few fields.
In this case, however, our metrics told us that the problem was pretty widespread. There wasn’t a particular slow request, rather, at what appeared to be random times, certain requests would be slow. We also realized that it wasn’t a particular request that was slow, but all requests within a given time period.
That hinted quite strongly at the problem, it was very likely that this is a GC issue that caused a long pause. We still had to do some work to rule out other factors, such as I/O / noisy neighbors, but we narrowed down on GC as being the root cause.
The problem in this case was that the customer was running multiple databases on the same RavenDB process. And each of them was fairly large. The total amount of memory that the RavenDB process was using was around 60GB of managed memory. At that level, anything that would cause a GC can cause a significant pause and impact operations.
The solution in this case was to break that into multiple separate processes, one for each database. In this manner, we didn’t have a single managed heap that the GC had to traverse, each heap was much smaller, and the GC pause times were both greatly reduced and spaced much further apart.
More posts in "Production postmortem" series:
- (07 Aug 2017) 30% boost with a single line change
- (04 Aug 2017) The case of 99.99% percentile
- (02 Aug 2017) The lightly loaded trashing server
- (23 Aug 2016) The insidious cost of managed memory
- (05 Feb 2016) A null reference in our abstraction
- (27 Jan 2016) The Razor Suicide
- (13 Nov 2015) The case of the “it is slow on that machine (only)”
- (21 Oct 2015) The case of the slow index rebuild
- (22 Sep 2015) The case of the Unicode Poo
- (03 Sep 2015) The industry at large
- (01 Sep 2015) The case of the lying configuration file
- (31 Aug 2015) The case of the memory eater and high load
- (14 Aug 2015) The case of the man in the middle
- (05 Aug 2015) Reading the errors
- (29 Jul 2015) The evil licensing code
- (23 Jul 2015) The case of the native memory leak
- (16 Jul 2015) The case of the intransigent new database
- (13 Jul 2015) The case of the hung over server
- (09 Jul 2015) The case of the infected cluster
Comments
Maybe this is a good testcase for Project Snowflake https://www.microsoft.com/en-us/research/publication/project-snowflake-non-blocking-safe-manual-memory-management-net/#
It would be interesting to hear what those gigs of RAM are composed of, in terms of both short- and long-lived objects. I know some of this might be private knowledge but I would imagine you're careful to use as much RAM as possible and use it to the greatest effect.
Do the various flavours of GC help at all with this? Server vs Workstation. Concurrent or not. I suspect you're using server and concurrent but am still curious if you've tried doing regular GC compaction anyway, and if you've tried asking it to do the large object heap or not at the same time.
Pop, We are looking at Snowflake very seriously. It is something that could really help us control what is going on, and that is GOOD.
Rik, We are assuming that the split processes are using roughly the same amount of memory, actually Or even a bit higher, probably. The savings here is that when a GC run, it has to do a lot less and doesn't impact everything at once.
Ian, We are already running in server mode, so I don't think it would matter too much. And we aren't compacting the LOH at all.
Comment preview
Join the conversation...