Feature #10181

Updated by Tom Clegg almost 2 years ago

Story: job output logged to keep while job is running

As a user, I would like to be able to retrieve complete (i.e. starting at the beginning and without being silenced by rate limiting) output of any running jobs. My strong preference would be to use keep for this, since (a) the full job log will be in keep anyway, so I'm already used to using that as an interface and (b) the bulk storage available to keep is generally going to be much greater than other places the logs could be kept (such as in the database). I would be ok with not necessarily always having up-to-the minute job logs available in keep, as long as the logs that are there are complete up to the point where they are truncated. Perhaps the final line of a truncated log entry could note that the job is still running and more output will arrive soon, along with stating the timestamp of the point at which the logs were flushed to keep (i.e. I would then know to expect that the next line would be timestamped after that time).

As a sysadmin, I would like to be able to adjust settings for flushing job logs to keep. I assume that any time a crunch job has a full block (i.e. 64MB) of output that it would be immediately written to keep and that the job's log collection would be updated to point to a new portable data hash which includes the new block. However, it would also be good to have a setting for flushing smaller amounts of log data to keep, so that logs from jobs that haven't output very much in some time can nonetheless be available. For example, I might configure a setting such that job output would be written to keep and the collection portable data hash updated every 15m regardless of how much output has been produced. That configuration option would be a tradeoff between creating a potentially large number of partially used keep blocks (although they would end up being cleaned up by keep-balance once a collection no longer points to them) and having a wait a long time for job output to appear in keep.

(remainder moved to #14284)

Story: job output does not belong in the database logs table and should be able to be directed to non-Arvados logging systems- systems

-As As a sysadmin, I'd rather my postgres database not fill up with hundreds of GB of job output logs. In addition to requiring a large amount of storage on the volume where the postgres database lives, this also tends to make queries to the logs table that have nothing to do with job output logging (i.e. fulfilling its role as more of an audit-log, such as checking for recent changes to collections) take ridiculously long. I think it would be best if no job output at all was stored in the central postgres database. In conjunction with the above story regarding storing in-progress job logs to keep, it would be great if some other system which is better suited to the task of buffering and distributing recent job output in order to make real-time job output available. It would be great if it could be sent via an existing log broker system such as logstash or fluentd such that it would be possible to not only direct the logs to whatever component Arvados uses to buffer and deliver the logs to consumers (such as via the existing websockets interface) but also to other non-Arvados logging systems (where we may be running the rest of the ELK/EFK stack for search and visualisation).-