[Gluster-devel] break glusterd into small parts (Re: good job on fixing heavy hitters in spurious regressions)

[Krishnan Parthasarathi]

> Why not break glusterd into small parts and distribute the load to
> different people? Did you guys plan anything for 4.0 for breaking glusterd?
> It is going to be a maintenance hell if we don't break it sooner.

Good idea. We have thought about it. Just re-architecting glusterd doesn't
(and will not) solve the division of responsibility issue that is being discussed here.
It's already difficult to maintain glusterd. I have already explained the reasons
in the previous thread.

> 
> Glusterd does a lot of things: Lets see how we can break things up one
> thing at a time. I would love to spend some quality time thinking about
> this problem once I am done with ec work, but this is a rough idea I
> have for glusterd.
> 
> 1) CLI handling:
> Glusterd-cli-xlator should act something like fuse in fs. It just gets
> the commands and passes it down, just like fuse gets the fops and passes
> it down. In glusterd process there should be snapshot.so, afr-cli.so,
> ec-cli.so, dht-cli.so loaded as management-xlators.
> Just like we have fops lets have mops (management operations).
> LOCK/STAGE/BRICK-OP/COMMIT-OP if there are more add them as well. Every
> time the top xlator in glusterd receives commands from cli, it converts
> the params into the arguments (req, op, dict etc) which are needed to
> carryout the cli. Now it winds the fop to all its children. One of the
> children is going to handle it locally, while the other child will send
> the cli to different glusterds that are in cluster. Second child of
> gluster-cli-xlator (give it a better name, but for now lets call it:
> mgmtcluster) will collate the responses and give the list of responses
> to glusterd-cli-xlator, it will call COLLATE mop on the first-child(lets
> call it local-handler) to collate the responses, i.e. logic for
> collating responses should also be in snapshot.so, afr-cli.so,
> dht-cli.so etc etc. Once the top translator does LOCK, STAGE, BRICK-OP,
> COMMIT-OP send response to CLI.
> 
> 2) Volinfo should become more like inode_t in fs where each *-cli xlator
> can store their own ctx like snapshot-cli can store all snapshot related
> info for that volume in that context and afr can store afr-related info
> in the ctx. Volinfo data strcuture should have very minimal information.
> Maybe name, bricks etc.
> 
> 3) Daemon handling:
>       Daemon-manager xlator should have MOPS like START/STOP/INFO and
> this xlator should be accessible for all the -cli xlators which want to
> do their own management of the daemons. i.e. ec-cli/afr-cli should do
> self-heal-daemon handling. dht should do rebalance process handling etc.
> to give an example:
> while winding START mop it has to specify the daemon as
> "self-heal-daemon" and give enough info etc.
> 
> 4) Peer handling:
>      mgmtcluster(second child of top-xlator) should have MOPS like
> PEER_ADD/PEER_DEL/PEER_UPDATE etc to do the needful. top xlator is going
> to wind these operations based on the peer-cli-commands to this xlator.
> 
> 5) volgen:
>      top xlator is going to wind MOP called GET_NODE_LINKS, which takes
> the type of volfile (i.e. mount/nfs/shd/brick etc) on which each *-cli
> will construct its node(s), stuff options and tell the parent xl-name to
> which it needs to be linked to. Top xlator is going to just link the
> nodes to construct the graph and does graph_print to generate the volfile.
> 
> I am pretty sure I forgot some more aspects of what glusterd does but
> you get the picture right? Break each aspect into different xlator and
> have MOPS to solve them.

We have some initial ideas on how glusterd for 4.0 would look like. We won't be
continuing with glusterd is also a translator model. The above model would
work well only if we stuck with the stack of translators approach.