Many thanks to @iverberk for the original PR (#1609), but we ended up
not wanting to ship this implementation with 0.5.
We'll come back to it after 0.5 and hopefully find a way to leverage
filesystem accounting and quotas, so we can skip the expensive polling.
* In -dev mode, `ioutil.TempDir` is used for the alloc and state
directories.
* `TempDir` uses `$TMPDIR`, which os OS X contains a per user
directory which is under `/var/folder`.
* `/var` is actually a symlink to `/private/var`
* Docker For Mac validates the directories that are passed to bind and on
OS X. That whitelist contains `/private`, but not `/var`. It does not
expand the path, and so any paths in `$TMPDIR` fail the whitelist check.
And thusly, by expanding the alloc/state directories the value passed
for binding does contain `/private` and Docker For Mac is happy.
* Added the alloc dir move
* Moving allocdirs when starting allocations
* Added the migrate flag to ephemeral disk
* Stopping migration if the allocation doesn't need migration any more
* Added the GetAllocDir method
* refactored code
* Added a test for alloc runner
* Incorporated review comments
Remove use of periodic consul handlers in the client and just use
goroutines. Consul Discovery is now triggered with a chan instead of
using a timer and deadline to trigger.
Once discovery is complete a chan is ticked so all goroutines waiting
for servers will run.
Should speed up bootstraping and recovery while decreasing spinning on
timers.
rpcproxy is refactored into serverlist which prioritizes good servers
over servers in a remote DC or who have had a failure.
Registration, heartbeating, and alloc status updating will retry faster
when new servers are discovered.
Consul discovery will be retried more quickly when no servers are
available (eg on startup or an outage).
Per discussion, we want to be aggressive about fanning out vs possibly
fixating on only local DCs. With RPC forwarding in place, a random walk
may be less optimal from a network latency perspective, but it is guaranteed
to eventually result in a converged state because all DCs are candidates
during the bootstrapping process.
Client: Search limit increased from 4 random DCs to 8 random DCs, plus nearest.
Server: Search factor increased from 3 to 5 times the bootstrap_expect.
This should allow for faster convergence in large environments (e.g.
sub-5min for 10K Consul DCs).
It's possible that a Nomad Client is heartbeating with a Nomad server that
has become issolated from the quorum of Nomad Servers. When 3x the
heartbeatTTL has been exceeded, append the Consul server list to the primary
primary server list. When the next RPCProxy rebalance occurs, there is a
chance one of the servers discovered from Consul will be in the majority.
When client reattaches to a Nomad Server in the majority, it will include
a heartbeat and will reset the TTLs *AND* will clear the primary server list
to include only values from the heartbeat.
This has been done to allow the Server and Client to reuse the same
Syncer because the Agent may be running Client, Server, or both
simultaneously and we only want one Syncer object alive in the agent.
I *KNEW* I should have done this when I wrote it, but didn't want to
go back and audit the handlers to include the appropriate return
handling, but now that the code is taking shape, make this change.
In addition to the API changing, consul.Syncer can now be signaled
to shutdown via the Shutdown() method, which will call the Run()'ing
sync task to exit gracefully.
While breaking the API within this PR, break out the individual
arguments to RefreshServerLists. The servers parameter is reusing
`structs.NodeServerInfo` for the time being, but this can be revisited
if the needs of the strucutre diverge in the future.
