Federation likes (votes) are wildly different from server to server as it stands now. And unless something is way off on my particular server, 0.4 seconds is what PostgreSQL is reporting as the mean (average) time per single comment vote INSERT, and post vote INSERT is similar. (NOTE: my server is classic hard drives, 100MB/sec bencharked, not a SSD)
Discussion of the SQL statement for a single comment vote insert: https://lemmy.ml/post/1446775
Every single VOTE is both a HTTP transaction from the remote server and a SQL transaction. I am looking into Postgress supporting batches of inserts to not check all the index constraints at each single insert: https://www.postgresql.org/docs/current/sql-set-constraints.html
Can the Rust code for inserts from federation be reasonably modified to BEGIN TRANSACTION only every 10th comment_like INSERT and then do a COMMIT of all of them at one time? and possibly a timer that if say 15 seconds passes with no new like entries from remote servers, do a COMMIT to flush based a timeout.
Storage I/O writing for votes alone is pretty large…
Or, build some sort of in memory buffer that bulks the actions, and commits to the DB in batches over time/size a-la AWS Kinesis Firehose.
Every action gets recorded in memory, in their own separate buffers (I.E. one for post, one for comment, one for like/dislike, etc.) and is supplemented to the results on each request. When the buffer reaches X MB in size, it is then batch insert/update’ed into the database as to reduce the overhead?
Well, I’m pulling out “like/dislike” (votes), because I consider it less of a priority. The actual comments and postings are taking over 1.0 second to INSERT, but that’s the bulk of the site’s purpose - to share actual content. If the likes lag by 15 seconds, is that such a big deal?
The way I’d imagine this working is the data is supplemented (augmented) into the query results.
Basically, there’d be a new cache / injection layer that sits between the application and the database. Instead of application directly working with the existing ORM to work with the DB (I’m not actually sure how Rust does this, so I’m just speaking in broader terms), the application would work against this layer that creates the buffer, and interface with the ORM. Then, on write actions, it fills up the buffer until buffer fills up or some time has past before bulk performing the write action; on read actions, it interfaces with ORM, and then weave the buffered data into the response back to the application.
Thus, from the user’s perspective, nothing should be changed, and they wouldn’t know any wiser.