[Plasma] Improve plasma documentation on distributed storage

[danlg]

What is your question?

From the documentation https://arrow.apache.org/blog/2017/08/08/plasma-in-memory-object-store/ and https://arrow.apache.org/docs/python/plasma.html#starting-the-plasma-store

it is not clear if Plasma storage is distributed on several nodes or not.

I assumed in the example /tmp/plasma if a local FS not an NFS mounted on several hosts.

Please advise.

Also it would be interesting to describe how Plasma compares to Memcached

[rkooo567]

Yes. We should definitely improve documentation about this.

About your question, here are answers.

• Is it distributed? Yes. It is distributed at each node (each plasma store is a part of raylet), and they don't use NFS. Objects are pulled and pushed to different nodes upon needs (I believe through RPC).

Plasma store is built to support zero-copy read. It stores objects in shared memory and use Apache Arrow to support zero-copy serialization among multiple processes

[danlg]

Hello @rkooo567 would you have any news on this one ?

[rkooo567]

@danlg Not much since I am a little oversubscribed lately. cc @suquark do you have some time to do this?

[harelwa]

Hi,

adding to this that the current Ray docs state:

Plasma is an in-memory object store that is being developed as part of Apache Arrow

This is misguiding, as ray "forked" from arrow due to reasons discussed here -

#7901

Even more confusing is that Arrow's Plasma will be deprecated in v10.0.0 -

https://arrow.apache.org/docs/python/plasma.html

I'd be happy to contribute a docs PR with the correction.

But, in case we'd like to add a "historical note" I'll need someone clarify the historical details ( @suquark ? ) -e.g. -

Who authored for the original Arrow Plasma Store ? ( clue here - https://lists.apache.org/thread/nw232k2lzmg9kcl8ts475m9ybl34j81p ) ?

Thanks,
Harel