When client nodes are restarted, all allocations that have been scheduled on the
node have their modify index updated, including terminal allocations. There are
several contributing factors:
* The `allocSync` method that updates the servers isn't gated on first contact
with the servers. This means that if a server updates the desired state while
the client is down, the `allocSync` races with the `Node.ClientGetAlloc`
RPC. This will typically result in the client updating the server with "running"
and then immediately thereafter "complete".
* The `allocSync` method unconditionally sends the `Node.UpdateAlloc` RPC even
if it's possible to assert that the server has definitely seen the client
state. The allocrunner may queue-up updates even if we gate sending them. So
then we end up with a race between the allocrunner updating its internal state
to overwrite the previous update and `allocSync` sending the bogus or duplicate
update.
This changeset adds tracking of server-acknowledged state to the
allocrunner. This state gets checked in the `allocSync` before adding the update
to the batch, and updated when `Node.UpdateAlloc` returns successfully. To
implement this we need to be able to equality-check the updates against the last
acknowledged state. We also need to add the last acknowledged state to the
client state DB, otherwise we'd drop unacknowledged updates across restarts.
The client restart test has been expanded to cover a variety of allocation
states, including allocs stopped before shutdown, allocs stopped by the server
while the client is down, and allocs that have been completely GC'd on the
server while the client is down. I've also bench tested scenarios where the task
workload is killed while the client is down, resulting in a failed restore.
Fixes#16381
When fetching node alloc assignments, be defensive against a stale read before
killing local nodes allocs.
The bug is when both client and servers are restarting and the client requests
the node allocation for the node, it might get stale data as server hasn't
finished applying all the restored raft transaction to store.
Consequently, client would kill and destroy the alloc locally, just to fetch it
again moments later when server store is up to date.
The bug can be reproduced quite reliably with single node setup (configured with
persistence). I suspect it's too edge-casey to occur in production cluster with
multiple servers, but we may need to examine leader failover scenarios more closely.
In this commit, we only remove and destroy allocs if the removal index is more
recent than the alloc index. This seems like a cheap resiliency fix we already
use for detecting alloc updates.
A more proper fix would be to ensure that a nomad server only serves
RPC calls when state store is fully restored or up to date in leadership
transition cases.
I noticed that `watchNodeUpdates()` almost immediately after
`registerAndHeartbeat()` calls `retryRegisterNode()`, well after 5
seconds.
This call is unnecessary and made debugging a bit harder. So here, we
ensure that we only re-register node for new node events, not for
initial registration.
Tons left to do and lots of churn:
1. No state saving
2. No shutdown or gc
3. Removed AR factory *for now*
4. Made all "Config" structs local to the package they configure
5. Added allocID to GC to avoid a lookup
Really hating how many things use *structs.Allocation. It's not bad
without state saving, but if AllocRunner starts updating its copy things
get racy fast.
This PR removes deepcopying of the job attached to the allocation in the
alloc runner. This operation is called very often so removing reflect
from the code path and the potentially large number of mallocs need to
create a job reduced memory and cpu pressure.
Tim Gross