[PUP-3647] Puppet stores files forever in state.yaml Created: 2014/11/07 Updated: 2018/10/02 Resolved: 2018/09/26
|Fix Version/s:||PUP 5.5.7, PUP 6.0.1|
|Reporter:||Peter Bakkum||Assignee:||Jacob Helwig|
|Remaining Estimate:||Not Specified|
|Time Spent:||Not Specified|
|Original Estimate:||Not Specified|
|Template:||PUP Bug Template customfield_10700 58064|
|Epic Link:||Puppet Performance|
|Sprint:||Platform Core KANBAN|
|Method Found:||Customer Feedback|
|CS Frequency:||3 - 25-50% of Customers|
|CS Severity:||3 - Serious|
|CS Business Value:||4 - $$$$$|
|CS Impact:||Many users would like to use recurse. Having it lead to poor performance isn't good.|
|Release Notes:||Bug Fix|
|Release Notes Summary:||Previously, the state.yaml file could grow unbounded. The new `statettl` setting controls how long entries are cached (default: 32 days). If you use resource schedules, please see the `statettl` documentation to see how this setting interacts with the schedule type.|
We noticed the following today on one of our production clusters: https://cloudup.com/cgJm3eD4wu4. Essentially Puppet Agent has been using gradually more CPU and Memory over time and was taking longer and longer to execute. On some nodes its been using 2G during each run.
Digging in, we discovered why. We have configuration like the following:
It turns out that the /var/db directory on this machine sometimes includes temporary files, because of the data system using this directory. This means there are a few new files in /var/db every time we run puppet agent.
We discovered today that each time puppet sees a file, it writes a line to /var/lib/puppet/state/state.yaml. Since there were always new files each time puppet ran, this file grew and grew. Its about 125MB on some of our machines. With a state.yaml file of that size, puppet agent took this long:
After deleting /var/lib/puppet/state/state.yaml it dropped to:
We've removed our use of `recurse` for now, which wasn't really necessary in this case, but we consider this a dangerous situation and a bug in Puppet; it means that we need to be sure Puppet isn't touching new files over time. Happy to provide more details if its helpful.
|Comment by Eric Sorenson [ 2017/06/26 ]|
Hm, I filed the linked duplicate bug
(If it turns out to actually be necessary, I'm not sure what the right approach would be to solve this problem specifically - seems like we'd need to put some pretty complicated logic in to prune a new state file of resources that used to be managed in (current_catalog - 1) but are no longer in (current_catalog)....)
This bit from lib/puppet/resource/catalog.rb appears to be the only usage of the statefile routines....
|Comment by Charlie Sharpsteen [ 2017/06/26 ]|
Looks like the ResourceHarness is the primary user of data stored in the statefile. Statefile data seems to come into play in two places:
Determining schedules. The :checked timestamp is used to determine whether a resource falls within a schedule:
Providing input for event generation. The cache is used to create historical_values:
Which are referenced in the event generation code:
I suspect the main use here may be to provide input for audit events.
A simple resolution to the growth of the file may be to enforce a time-to-live when the state is flushed out to disk:
For example: filtering out all entries that are older than some multiple of the agent runinterval.
|Comment by Josh Cooper [ 2017/06/26 ]|
Also copying a comment from Michael Smith from the duplicate ticket, we load the state.yaml file twice.
|Comment by Henrik Lindberg [ 2018/03/14 ]|
I started to read code to figure out what is going on - beginning with why the state file is loaded twice. The two loads (in the order they occur) come from:
That raises the question - why is catalog doing this only if "host_config?" and why is it doing that? Is it a valid reload because it should start from the original stored state? Can someone explain what "host_config?" is?
A bit further down in catalog.rb we can read this:
So, clearly the cache is never written by an apply unless it is a "host_config?"
The load caused by the configurer seems to be because it uses the cache, and attempt to rescue a bad situation:
The line that sets @compile_time seems to be why it needs to read it from loaded state. Anyone know what this is about?
The code seems to be from 2007 written by Luke.
I think the two loads comes from first applying the settings catalog, and then applying the real catalog. I base this on seeing catalog.apply being called three times, and the first time host_config? is false, the second time it is true, third time it is false. It is a guess though. I also wonder what the third apply is for?
The handling of errors from the load that is done in configurer looks wrong - there is similar handling in Catalog that seems to do a better job - renaming the file with a "bad" so it can be examined, and also handling the issue of not being able to remove (say on windows).
Interestingly - if I brute force just return in configurer init_storage to this:
It still passes all tests in the puppet repo. (Which makes me sad...)
|Comment by Charlie Sharpsteen [ 2018/03/14 ]|
The biggiest effect of checking host_config? should be to avoid loading the state when applying the "settings catalogs" that Puppet uses to create cache directories and manage other files related to settings in puppet.conf, or to download plugins by applying a recursive file resource:
The whole reasoning behind the load in init_storage seems to be to retrieve the "compile_time", which appears to never be used, and to clean the state file out if it happens to be corrupted. If we're not using the compile_time, or interacting with the state in Puppet::Util::Storage until after a catalog has been retrieved from the Puppet Server (or loaded from cache), then it seems like we can get rid of the init_storage call in the configurer.
As for the state file its self, adding a time to live on the entries that is used to prune data when we write the file back at the end of a run seems like a good way to get rid of the bloat problem.
|Comment by Henrik Lindberg [ 2018/03/15 ]|
Charlie Sharpsteen Thanks for additional input. See if someone has any more to say about @compile_time.
For the second part of the problem - if it prunes based on write time it needs to touch all entries that are used - otherwise the cache would not have entries that originated from 2 x run-interval time ago. It would also mean that if a sequence of catalogs are used with some smaller intermediate catalogs then those would result in cache evictions. Is that acceptable?
|Comment by Charlie Sharpsteen [ 2018/03/15 ]|
I'm assuming that if a resource is in the catalog or generated agent-side as part of something like a recursive file resource, then it gets written to state.yaml with an updated :checked timestamp. If this is the case, then we could apply a simple filter during Puppet::Util::Storage.store that drops all entries where :checked is older than some time-to-live:
For :statettl, we could probably just pick something like "7 days" as a default value and solve 80% of the problem. Things that generate a lot of catalog entries, like recursive file resources that return 10s or 100s of thousands of directory entries, will still be a problem. But, these sorts of resources should be avoided anyway as they cause performance problems beyond the size of the statefile.
|Comment by C Lang [ 2018/08/15 ]|
We were just bitten by this one in a scenario where Tidy was used to purge a handful of temp files every run - after 6 months, the state file was 70M and the agent hammered the CPU for 2 minutes before doing anything.
I'm surprised this isn't a higher priority after four years. If it can't be fixed, maybe Tidy should be removed or at least have a warning that it will eventually bring the agent to its knees?
|Comment by Stefan Strandberg [ 2018/08/30 ]|
This bit us where we were briefly using a Tidy to clear puppet reports older than 10 days on our puppetserver. We realized that the Tidy was causing issues several months ago and changed it to just a cron job, which resolved an immediate issue, but puppet client runs on the puppetserver were taking >5 minutes still. I finally spent time today diagnosing and ultimately ran into this thread and discovered that the state file was over 300MB despite the problematic Tidy being taken out a very long time ago. After removing it, puppet runs are back down to sub-6 seconds of actual time, 16 seconds total. Puppetserver load which had been pinned around 3-5 dropped immediately to <0.5 as well, improving performance across the board.
I'd like to suggest a higher priority for this as well because it's not inherently obvious that there is a state file, that it tracks every file puppet was ever aware of, and that it doesn't stop tracking files that it should no longer care about, leading it to grow forever and ever until it bogs down the system. Our planned "workaround" is going to be a cron job that kills it once a month everywhere, but I worry about unintended consequences with that and would much prefer an internal solution as described above.
|Comment by Jarret Lavallee [ 2018/09/19 ]|
The Schedule resource uses the state file to determine if the schedule matches, so a short ttl could break longer schedules. If we implement a :statettl, please be sure to update the Schedule documentation to denote that it needs to be configured for a longer time. Alternatively, we could check to see if the resource has a schedule while pruning and skip those items.
|Comment by Kris Bosland [ 2018/09/24 ]|
Merged into 5.5.x at b4901b3.
|Comment by Josh Cooper [ 2018/09/26 ]|
Passed CI in e2f7ccfb7