Production postmortemThe case of 99.99% percentile
time to read 2 min | 326 words

imageA 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.

Tweet Share Share 6 comments
Tags:

Related posts that you may find interesting:

Postmortem: The case of the pauses, the fault and the GC When your memory won’t go down What is my query doing?

More posts in "Production postmortem" series:

  1. (07 Apr 2025) The race condition in the interlock
  2. (12 Dec 2023) The Spawn of Denial of Service
  3. (24 Jul 2023) The dog ate my request
  4. (03 Jul 2023) ENOMEM when trying to free memory
  5. (27 Jan 2023) The server ate all my memory
  6. (23 Jan 2023) The big server that couldn’t handle the load
  7. (16 Jan 2023) The heisenbug server
  8. (03 Oct 2022) Do you trust this server?
  9. (15 Sep 2022) The missed indexing reference
  10. (05 Aug 2022) The allocating query
  11. (22 Jul 2022) Efficiency all the way to Out of Memory error
  12. (18 Jul 2022) Broken networks and compressed streams
  13. (13 Jul 2022) Your math is wrong, recursion doesn’t work this way
  14. (12 Jul 2022) The data corruption in the node.js stack
  15. (11 Jul 2022) Out of memory on a clear sky
  16. (29 Apr 2022) Deduplicating replication speed
  17. (25 Apr 2022) The network latency and the I/O spikes
  18. (22 Apr 2022) The encrypted database that was too big to replicate
  19. (20 Apr 2022) Misleading security and other production snafus
  20. (03 Jan 2022) An error on the first act will lead to data corruption on the second act…
  21. (13 Dec 2021) The memory leak that only happened on Linux
  22. (17 Sep 2021) The Guinness record for page faults & high CPU
  23. (07 Jan 2021) The file system limitation
  24. (23 Mar 2020) high CPU when there is little work to be done
  25. (21 Feb 2020) The self signed certificate that couldn’t
  26. (31 Jan 2020) The slow slowdown of large systems
  27. (07 Jun 2019) Printer out of paper and the RavenDB hang
  28. (18 Feb 2019) This data corruption bug requires 3 simultaneous race conditions
  29. (25 Dec 2018) Handled errors and the curse of recursive error handling
  30. (23 Nov 2018) The ARM is killing me
  31. (22 Feb 2018) The unavailable Linux server
  32. (06 Dec 2017) data corruption, a view from INSIDE the sausage
  33. (01 Dec 2017) The random high CPU
  34. (07 Aug 2017) 30% boost with a single line change
  35. (04 Aug 2017) The case of 99.99% percentile
  36. (02 Aug 2017) The lightly loaded trashing server
  37. (23 Aug 2016) The insidious cost of managed memory
  38. (05 Feb 2016) A null reference in our abstraction
  39. (27 Jan 2016) The Razor Suicide
  40. (13 Nov 2015) The case of the “it is slow on that machine (only)”
  41. (21 Oct 2015) The case of the slow index rebuild
  42. (22 Sep 2015) The case of the Unicode Poo
  43. (03 Sep 2015) The industry at large
  44. (01 Sep 2015) The case of the lying configuration file
  45. (31 Aug 2015) The case of the memory eater and high load
  46. (14 Aug 2015) The case of the man in the middle
  47. (05 Aug 2015) Reading the errors
  48. (29 Jul 2015) The evil licensing code
  49. (23 Jul 2015) The case of the native memory leak
  50. (16 Jul 2015) The case of the intransigent new database
  51. (13 Jul 2015) The case of the hung over server
  52. (09 Jul 2015) The case of the infected cluster