Support ephemeral nodes

[ha]

Support ephemeral nodes

[dgryski]

ActiveState's fork ( https://github.com/ActiveState/doozerd ) has support for ephemeral nodes.

[pitr]

awesome. let's keep this issue open until it's merged and documented

[bketelsen]

Patches GLADLY reviewed.

[mreiferson]

I poked around ActiveState's fork and, to be honest, I'm not terribly happy with the implementation chosen.

I think any node should be able to be created as ephemeral (not just some specific hard coded path requiring special case handling in both get and set code).

I think this would produce the most flexible implementation but I acknowledge that it might mean more breaking changes to the API.

I'm curious to hear other's thoughts before I take a swing at this.

[pitr]

@mreiferson I also looked at the code and I agree. It's obvious that ActiveState did what anyone else would have done in their place - quickly get the feature to work. But for it to be useful to everyone, this needs to be done properly and thoughtfully. Before we hit major version 1 I think it's perfectly acceptable to break API, considering how important this feature is.

[mreiferson]

@pitr Yes, absolutely, it's always a matter of tradeoffs 😄

[srid]

It's obvious that ActiveState did what anyone else would have done in their place - quickly get the feature to work.

that's exactly what happened (the commit itself was titled "limited ephemeral node support").

I think any node should be able to be created as ephemeral (not just some specific hard coded path requiring special case handling in both get and set code).

+1

But for it to be useful to everyone, this needs to be done properly and thoughtfully. Before we hit major version 1 I think it's perfectly acceptable to break API, considering how important this feature is.

as for the API, how about having SET take an extra argument (t.req.CreateMode = EPHEMERAL)?

as for the implementation, i'm thinking that the list of ephemeral paths can be stored in the conn{} struct; to be deleted during disconnection.

[mreiferson]

@srid I was thinking the same thing (your suggested implementation)

Do you want to take a stab at this? I'm working on cleaning up logging and other misc odds and ends so I might not get to this for a bit.

[srid]

sure, i'll take a look.

[pitr]

should directories have ability to be ephemeral?

[mreiferson]

@pitr at the moment you only ever indirectly create directories (SET works on "file" paths), so my vote would be no.

[bketelsen]

Agreed. And there's no real need for data-owning directories in the use cases I've seen.

[cespare]

I've certainly seen Zookeeper use cases where ephemeral znodes with other nodes underneath them (ephemeral directories in Doozer) are useful. For instance, a server might come online, open up an ephemeral node called $HOSTNAME to indicate its existence and keep status information in files underneath that.

But since you don't really deal with directories directly in Doozer, I see how that wouldn't really make sense here.

[bketelsen]

Sorry my comment is unclear. I meant data-owning in the sense that all nodes in zk can have data associated, but they can't in doozer. A path like /myHostname/ can't have data attached, but an easy work around is to apply children paths like /myHostname/ipAddress with the data you want to store.