Ephemeral node deleted by connection loss prevents parent node deletion

[tsuraan]

This one's a bit weird, but basically if you create some node, and then create a sequential node within it, the node goes away on client connection loss, but the parent cannot be deleted because it thinks it isn't empty. Here's a sample (using the python kazoo library):

Basic setup:

Python 3.4.5 (default, Jan 26 2017, 00:57:26) 
Type 'copyright', 'credits' or 'license' for more information
IPython 6.2.1 -- An enhanced Interactive Python. Type '?' for help.
In [1]: from kazoo.client import KazooClient
In [2]: c=KazooClient()
In [3]: c.start()
In [4]: c.get_children('/')
Out[4]: ['']

Okay, now we can create foo, create foo/bar as ephemeral, fail to delete foo as expected, delete foo/bar, and then deleting /foo works also as expected:

In [5]: c.create('/foo')
Out[5]: '/foo'
In [6]: c.create('/foo/bar', ephemeral=True)
Out[6]: '/foo/bar'
In [7]: c.get_children('/foo')
Out[7]: ['bar']
In [8]: c.delete('/foo')
---------------------------------------------------------------------------
NotEmptyError                             Traceback (most recent call last)
...
NotEmptyError: 

In [9]: c.delete('/foo/bar')
Out[9]: True

In [10]: c.delete('/foo')
Out[10]: True

In [11]: 

Now for the weird; create /foo as before, create /foo/bar ephemeral as before, but disconnect and reconnect the client, then try to delete /foo (which appears to be empty):

In [11]: c.create('/foo')
Out[11]: '/foo'

In [12]: c.create('/foo/bar', ephemeral=True)
Out[12]: '/foo/bar'

In [13]: c.stop()

In [14]: c.close()

In [15]: c.start()

In [16]: c.get_children('/foo')
Out[16]: []

In [17]: c.delete('/foo')
---------------------------------------------------------------------------
NotEmptyError                             Traceback (most recent call last)
...
NotEmptyError: 

Well, that was weird. /foo looks empty, but deleting it says it isn't empty. Can we fix things? Yes, re-creating /foo/bar and then explicitly deleting it lets us delete /foo again:

In [18]: c.create('/foo/bar')
Out[18]: '/foo/bar'

In [19]: c.delete('/foo/bar')
Out[19]: True

In [20]: c.delete('/foo')
Out[20]: True

So, that's weird.

[tsuraan]

I've been digging a bit, but I think the issue may be that the correlated etcd node's lifetime is tied to the lease of the entire zetcd server, and not to any particular client's connection. I haven't yet found why the ephemeral node disappears from zetcd at all, but I would guess that there's a disconnect between what zetcd is showing and what etcd has stored. I'll keep poking around.

[heyitsanthony]

Nope, leases are per-connection. I think I found an unrelated ephemeral node bug, though...

[tsuraan]

Sorry for the silence, that fix does work for me.

[andrewhao888]

Hello @heyitsanthony, I have met the same issue, and your fix commit also work for me I am just wondering when it will be merged into master? Thanks

[heyitsanthony]

/cc @gyuho

[gyuho]

Just merged. Thanks!