Production postmortemThe case of the lying configuration file
While studying an issue using customer data, I noticed that indexing speed wasn’t up to what I expected it to be. In fact, the size of the indexing batch remained roughly constant (and small), and didn’t exhibit the usual increases as RavenDB notices that the server has a lot of work to do and the resources to do it. This was while investigating something else, but since I had to re-index that database quite a few time, I decided to investigate what was going on.
The underlying issue turned out to be a configuration setup. An index was specified with a MaxNumberOfOutputsPerDocument of 55. We use this value for a few things, among them to ensure to manage the memory resulting from indexing operations. In particular, we have had some issues with indexes that output a large number of index entries per documents using more then the quota allocated to the index and generating (sometime severe) memory pressure.
Unfortunately, in this case, we had the other option. The index was configured properly, but the index didn’t actually output multiple entries. So we ended up assuming that the index would generate a lot more memory than it would actually really use. That meant that we couldn’t feed it larger batches, because we feared it would use too much memory…
The fix was to make effectively ignore this value. Instead of using the value assuming that the user knows what is going it, we’ll use this value as the maximum value only, and use heuristics to figure out how much memory we should reserve for the index in question.
This is a smaller example of a wider issue. The values that the system gets are user input. And they should be treated as such. This means that you need to validate them in the same sense you would validate any other input from users.
In the case outlined above, the only implication was that we would index a more slowly, and weren’t able to take full advantage of the machine resources we had available. There have been other such issues.
An administrator has setup a set of range values so the min value was larger than the max value. This turned out to cause us to have no effective limit, short of an integer overflow. The end result was that we would use unbounded memory for our needs, which would work, most of the time, except when you had a big set of changes to apply and we would try to do it all at once…
Another case was an ops guy that wanted to reduce RavenDB CPU usage, so he set the number of threads available for background work to 0. That meant that work that was queued on background threads would never complete, and the system would effectively hang.
You get the drift, I assume.
Your configuration file (or however you are actually configuring things) is yet another way for your user to communicate with your application. Sure, the tone is much stricter (commanding, rather than asking), but you still need to make sure that what the user is asking you to do actually make sense. And if it doesn’t, you need to decide what to do about it.
- You can exit the process. The “I can’t work with you people” mode of error handling.
- You can emit a warning and proceed. The “I told you so, buster!” motto of blame shifting.
- You can ignore the set value. The “I know better than you” school of thought.
There aren’t really good choices, especially for server applications that can’t just show an error to the user.
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