Our monitoring system pinged us about a problem with a RavenDB cluster running in production. The problem was simple, we saw quite a bit of server restarts for that particular cluster. Looking deeper, it was obvious that the RavenDB instances for the cluster would occasionally run out of memory and crash. The customer, by the way, was unaware of this issue. From their perspective, the RavenDB cluster would switch the primary node for the database in question on a regular basis. On our end, we could see that each node would start using higher and higher memory and end up dying because of that. They would be restarted, of course, and the primacy of the cluster would switch automatically, but that is not a proper way to run.
The problem was figuring out what was going on. It took some time to figure out what exactly was going on. We didn’t see any such behavior on any other customer, but this customer had two factors that affected the outcome. The first is that the database in question is encrypted, which means that RavenDB will need some place to put the decrypted values. The second is that the user is issuing streaming queries that have a lot of results. We were able to reproduce the high memory usage when issuing the same queries, however, we were utterly unable to reproduce the problem when trying to run it on our own machines.
That was… strange, and it took a while to figure out that we need to run on Linux to get the issue. We subjected the system to a very high load on Windows, with no issue. On Linux, it would be quickly apparent that we are consuming more and more memory. We were able to narrow things down to this call:
posix_memalign(&ptr, 4096, 8192);
What we are asking here is an 8KB buffer aligned on 4KB boundary. And we were leaking those like crazy but we couldn’t figure out how. We are pretty careful with manual memory management and we have the tools around to detect leaks. Each and every call to allocate was also freed. The problem is that we aren’t the only ones using the system. Basically, posix_memalign will use the same memory pool as malloc(). The problem is memory fragmentation, basically. The way posix_memalign() works is to issue:
Where nb is 8192 bytes, alignment is 4096 bytes and MINSIZE is 32 bytes. We then release the end of the buffer, which ends up being ~4KB or so in most cases. Along with other allocations, that created severe fragmentation in our memory.
We need the memory to be page aligned, because we use that for direct memory access. The memory is typically pooled, so we won’t be allocating and freeing it all the time, but when you use streaming queries, you may be running through a lot of data, so we exceeded the size of the pool. At that point, we would allocate (and free), but we’ll also fragment the memory.
We fixed the issue by using mmap() directly, which will give us page aligned memory and won’t cause us to use more memory than needed. Given that we get page aligned memory with is a multiple of page size, we can be sure that we’ll get reuse of the memory, instead of having to deal with internal fragmentation inside the malloc implementation. With this change, there are no issues, and we are actually slightly faster than before.
The reason we didn’t run into the same problem on Windows, by the way? There we called VirtualAlloc() from the get-go, which will ensure that we have page aligned memory, so no need to deal with fragmentation.
More posts in "Production postmortem" series:
- (27 Jan 2023) The server ate all my memory
- (23 Jan 2023) The big server that couldn’t handle the load
- (16 Jan 2023) The heisenbug server
- (03 Oct 2022) Do you trust this server?
- (15 Sep 2022) The missed indexing reference
- (05 Aug 2022) The allocating query
- (22 Jul 2022) Efficiency all the way to Out of Memory error
- (18 Jul 2022) Broken networks and compressed streams
- (13 Jul 2022) Your math is wrong, recursion doesn’t work this way
- (12 Jul 2022) The data corruption in the node.js stack
- (11 Jul 2022) Out of memory on a clear sky
- (29 Apr 2022) Deduplicating replication speed
- (25 Apr 2022) The network latency and the I/O spikes
- (22 Apr 2022) The encrypted database that was too big to replicate
- (20 Apr 2022) Misleading security and other production snafus
- (03 Jan 2022) An error on the first act will lead to data corruption on the second act…
- (13 Dec 2021) The memory leak that only happened on Linux
- (17 Sep 2021) The Guinness record for page faults & high CPU
- (07 Jan 2021) The file system limitation
- (23 Mar 2020) high CPU when there is little work to be done
- (21 Feb 2020) The self signed certificate that couldn’t
- (31 Jan 2020) The slow slowdown of large systems
- (07 Jun 2019) Printer out of paper and the RavenDB hang
- (18 Feb 2019) This data corruption bug requires 3 simultaneous race conditions
- (25 Dec 2018) Handled errors and the curse of recursive error handling
- (23 Nov 2018) The ARM is killing me
- (22 Feb 2018) The unavailable Linux server
- (06 Dec 2017) data corruption, a view from INSIDE the sausage
- (01 Dec 2017) The random high CPU
- (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