search

typhon-project / typhon-polystore-api

Typhon Polystore API
Eclipse Public License 2.0
2 stars 0 forks source link

State of Polystore components #14

Open MarieSaphira opened 4 years ago

MarieSaphira commented 4 years ago

Hi, to create deployment scrips for Kubernetes I need to know if the polystore components are stateless (except for the databases). WishedStates Can you confirm this? @DavyLandman @tvdstorm @pinoval @OrfenCLMS @zolotas4 @kolovos @Danny2097 @malwash

DavyLandman commented 4 years ago

Yup, looks about right.

Just for clarity, currently QL server is not stateless, but after fixing typhon-project/typhonql#68 we can be stateless.

MarieSaphira commented 3 years ago

Analytics and NLAE are not stateless. How about the API @Stratidakos ?

Stratidakos commented 3 years ago

The API is stateless.

DavyLandman commented 3 years ago

Hi @Stratidakos last time we talked it was statefull, since it contains keeps around the current model in memory? (as in, it's not re-checking the Polystore DB for updated models)

This is primarily a thing for when you run the API behind a load balancer. Then the update model call will only reach one of the instances.

If this is solved, cool 👍

Stratidakos commented 3 years ago

No, thank you for reminding me. The fact, that we keep the models in a MongoDB, makes the API stateful. We will change a few things to make it stateless.

theofilis commented 3 years ago

@DavyLandman I have a question just to clarify a thing on the @MarieSaphira's diagram?

The PolystoreDB is referring to storage where API stores its configurations or I misunderstood it?

MarieSaphira commented 3 years ago

yes, it's polystore-mongo

theofilis commented 3 years ago

Ok, this is great because. We found that the service is storing the model there. Give us some time to verify it.

If this is true, we can assume that the Polystore API is stateless because it stores its state to the polystore-mongo database.

Otherwise, we will modify it to store the model their.

How is it sound?

MarieSaphira commented 3 years ago

The DL and ML models are uploaded to the polystore-mongo at startup. If the API now only uses the models from the database and does not keep them in memory as it was before (see @DavyLandman post above) you are fine.

DavyLandman commented 3 years ago

Does that mean that everytime someone sends a query to the polystore api, it retrieves the models fresh from the mongo db?

theofilis commented 3 years ago

@DavyLandman good point, I'll check for this.

I supposed to make the API stateless, it should every time request the model from the Database [least effort], of course, this has a negative impact on query performance.

I am thinking, two possible solutions to minimize the impact, which are:

  1. To set-up the API to keep the model in memory and periodically request the version from the database.
  2. [most difficult] Or to set up a mechanism to update each instance of the Polystore API cached model.
DavyLandman commented 3 years ago

A good thing to consider is that the evolution tool will send model updates and queries for the new model quickly afterwards, so any caching has to take that into consideration.

Is there maybe a kind of mongo event you can subscribe to?

theofilis commented 3 years ago

@DavyLandman Ok I have never used MongoDB events. I have to study them a little.

DavyLandman commented 3 years ago

There appears to be a change stream: https://docs.mongodb.com/manual/changeStreams/ I've never used it, but it might be nice for this purpose?