No document can be fed because ostensibly there isn't enough disk space about pyvespa

[kkraune]

I have updated https://pyvespa.readthedocs.io/en/latest/troubleshooting.html#full-disk, linking to a new example in https://pyvespa.readthedocs.io/en/latest/application-packages.html , where one can set a higher limit.

I know this is a bit cumbersome, please give this a try and let me know.

https://docs.vespa.ai/en/proton.html#proton-maintenance-jobs means that Vespa needs disk space for compaction jobs. How much is schema-dependent, 75% is a conservative number - this helps operators avoiding index corruption due to full disk.

[bratseth]

1. When running locally I suggest configure disk size with Docker Desktop. Yes, Vespa could have its own way of setting disk limits, but it doesn't seem necessary given that Docker/Linux containers already do this, and running on a container is preferable in any case.
2. Since Vespa assumes it controls the entire disk given to it, it follows that if the disk is 90% full Vespa has already used 90% of it and should not use even more.

[baldersheim]

@ch3rn0v
1 - For development I see that I am running with a volume of 170G and using docker desktop for mac. In production I think we are using xfs anyway. Not a docker expert so I have no expert advise to give here.

2 - It is slightly more complicated as it is a variable number which initially will be very small and not enough to cater for the fixed overhead. Doing that would also allow other services using disk/memory to run on the same node, which would be a feature for development nodes, but make running a production system much more complicated. And since there are other ways to solve this issue we have not prioritized making our own solution.

[ch3rn0v]

Thanks a lot for your effort! If I were making design choices regarding the architecture I would fix the issue at its root. Which I believe is the fact that the disk space required by Vespa for "compaction jobs" is proportional to the space already occupied by the storage. It has absolutely nothing to do with the total disk space. An empty storage doesn't need free extra 100Gb to add ten 5Kb documents to an empty storage.

I wonder what other convenient choices were made by Vespa authors, but I'd rather not spend my time satisfying my curiosity so long as there are properly designed alternatives that just work.

[baldersheim]

Some time ago, in a distant past, the conclusion was vespa was made for large systems spanning multiple distributed machines focusing on scalability and ease of operation. From this the conclusion followed that apart for monitoring and other minor services vespa would be the only service running.

This still holds true.

If you want to present a smaller portion of your machine to vespa docker/podman is the better solution for that.

[ch3rn0v]

@baldersheim , thank you for the explanation! If you don't mind:

1. How does one experiment with Vespa locally or develop it before deploying to multiple distributed machines?
2. How is a proportion of already occupied space worse than a proportion of all available space?

I tried using docker. However, specifying "overlay2.size=10G" doesn't work for file systems other than xfs and passing --storage-opt size=10G to a docker call requires to either change the call in the Vespa's source code or pass it as an argument or config somewhere, but I don't see it being mentioned in Vespa's documentation anywhere.

[ch3rn0v]

Now that I see the reasoning behind this choice, I understand it much better. Thanks a lot for clarifying this! Unfortunately I didn't find any easy and clean way to run and test Vespa locally w/o Docker GUI, so perhaps I'll try it out some other time out of curiosity. For now running Weaviate seemed simple enough, but Vespa and Faiss is something I'd like to do in order to compare things.