This reverts commit 65d14ccd92.
This was actually a bad idea. For inexplicable reasons, running these
Ceph commands manually (not even via Python, but in a normal shell)
takes 7 * two orders of magnitude longer than running them with the
Rados module, so long in fact that some basic commands like "ceph
health" would sometimes take longer than the 1 second timeout to
complete. The Rados commands would however take about 1ms instead.
Despite the occasional issues when monitors drop out, the Rados module
is clearly far superior to the shell commands for any moderately-loaded
Ceph cluster. We can look into solving timeouts another way (perhaps
with Processes instead of Threads) at a later time.
Rados module "ceph health":
b'{"checks":{},"status":"HEALTH_OK"}'
0.001204 (s)
b'{"checks":{},"status":"HEALTH_OK"}'
0.001258 (s)
Command "ceph health":
joshua@hv1.c.bonilan.net ~ $ time ceph health >/dev/null
real 0m0.772s
user 0m0.707s
sys 0m0.046s
joshua@hv1.c.bonilan.net ~ $ time ceph health >/dev/null
real 0m0.796s
user 0m0.728s
sys 0m0.054s
Prevents bad states where the VM is "removed" but some of its disks
remain due to e.g. stuck watchers.
Rearrange the sequence so it goes stop, delete disks, then delete VM,
and then return a failure if any of the disk(s) fail to remove, allowing
the task to be rerun after fixing the problem.
Regenerating the ZK connection was fraught with issues, including
duplicate connections, strange failures to reconnect, and various other
wonkiness.
Instead let Kazoo handle states sensibly. Kazoo moves to SUSPENDED state
when it loses connectivity, and stays there indefinitely (based on
cursory tests). And Kazoo seems to always resume from this just fine on
its own. Thus all that hackery did nothing but complicate reconnection.
This therefore turns the listener into a purely informational function,
providing logs of when/why it failed, and we also add some additional
output messages during initial connection and final disconnection.
Using the Rados module was very problematic, specifically because it had
no sensible timeout parameters and thus would hang for many seconds.
This has poor implications since it blocks further keepalives.
Instead, remove the Rados usage entirely and go back completely to using
manual OS commands to gather this information. While this may cause PID
exhaustion more quickly it's worthwhile to avoid failure scenarios when
Ceph stats time out.
Closes#137
When trying to write to sub-item paths that don't yet exist, the
previous method would just blindly write to whatever the root key is,
which is never what we actually want.
Instead, check explicitly for a "base path" situation, and handle that.
Then, if we try to get a subpath that isn't valid, return None. Finally
in the various functions, if the path is None, just continue (or return
false/None) and (try to) chug along.
Like the other Celery job this does not work properly with the
ZKConnection decorator due to conflicting "self", so just connect
manually exactly like the provisioner task does.
Matches the new node list output format with the additional header line,
as well as revamps some other aspects:
1. Adjusts the UUID to be under the name in the info output.
2. Removes the UUID from the list output to save space, because this
is generally not needed in day-to-day quick-list output.
3. Renames the "Node" header to "Current" to better reflect what
that column actually means and avoid conflicting with the parent
header.
Also adjusts the layout of the node list output to avoid excessively
long lines. Adds another header line with categories and spacing dashes
for easier visual parsing.
Not sure how this didn't cause an issue until now, but the wrong key
path was used and this was getting unexpected data with the newly-added
version string instead of the proper mode string.
The zkhandler read() function takes care of ensuring there is a None
value returned if these fail, so these aren't required. Makes the code a
fair bit more readable here.
When doing a stop_vm or terminate_vm, check again after 0.2 seconds
and try re-terminating if it's still running. Covers cases where a VM
doesn't stop if given the 'stop' state.
This helps parallelize the numerous Zookeeper calls a little bit, at
least within the bounds of the GIL, to improve performance when getting
a large list of VMs. The max_workers value is capped at 32 to avoid
causing too many threads during concurrent executions, but still
provides a noticeable speedup (on the order of 0.2-0.4 seconds with 75
VMs, scaling up further as counts grow).
Usable anywhere that the global daemon "config" parameter can be passed
in (e.g. pvcapid/helper.py, pvcnoded/Daemon.py, etc.). Stores results in
a subdirectory of the PVC logdir called "profiler" if this directory can
be created, or prints results.
The debug config parameter ensures that the profiler can be added to
functions and not run unless the server is explicitly in debug mode.
Might not be useful as I don't initially plan to add this to every
function (only when investigating performance problems), but this
flexibility allows that to change later.
Instead of looping 5+ times acquiring an impossible lock on a
nonexistent key, just fail on a different error and return failure
immediately.
This is likely a major corner case that shouldn't happen, but better to
be safe than 500.
These cause a major (2x) slowdown in read calls since Zookeeper
connections are expensive/slow. Instead, just try the thing and return
None if there's no key there.
Also wrap the children command in similar error handling since that did
not exist and could likely cause some bugs at some point.