seeing more of this broken pipe and socket timeout errors which we had noticed before ,
https://discord.com/channels/907569970500743200/928649303898075196/1059893176082645043
the only thing which we are having diff is the efs , and not storing data in ebs.

( is anyone else has noticed this errors) now i am seeing at a retrofit calls as well at read time , and happening at read time makes it critical.

sharing the stacktrace at readtime

Hi, could you please provide the details of the request you send to Qdrant? Is there a chance you're setting the timeout on your end?

qdrant.connectTimeout=6000
qdrant.readTimeout=20000

its in ms. but yeah we do have timeout setup. this is for hitting recommend api from java app as a retrofit call .( for whihc i have shared the stacktrace )

and for writing data to qdrant in python app , its 300 . and its set like this. ( qdrant client version is 0.11.5 )

self.client = QdrantClient(host=app.config['VECTOR_DB_HOST'], port=app.config['VECTOR_DB_PORT'], timeout=Timeout(timeout=app.config['VECTOR_DB_TIMEOUT']))

And how about the recommendation API call? How many negative/positive examples do you provide? Is there any additonal filtering done? How many results do you expect?

The bottleneck could also eventually be EFS cause it's unable to provide the low disk latency per IO operation like EBS.

Yeah 💯

Sorry , for late reply its only one point in positive .I also suspect efs only 😅 but my concern is how can i prove that as if i suggest that as solution ,and issue still persost

Also efs have a benifit of sharing and if its easier to take a backup using aws backup with but with ebs backup would be only s3 and would udeally want to take data backup daily.

You can take ebs backups also daily and even incrementally. I haven't worked much with efs, but it is not meant to be used for workloads like databases and I am not 100% sure if the backups would even be consistent. Ebs Snapshots are convenient to do and can also easily be transferred and recovered.

Proving that efs is the bottleneck is a bit tricky, but you could try and take a look at io_wait for example