CI systems sometimes hiccup for five seconds, which causes heartbeat leases
to fail and breaks all sorts of things. As a workaround, update heartbeat
TTL for tests only. This might in turn cause different failures to do
with leader elections, but in theory those should be handled now?
A while back we started just polling the reported cluster state of one node
when starting a cluster for tests. This works fine if we're doing fresh
new etcd queries for every single operation -- but that's insanely
expensive, it turns out.
When we use the watcher, some nodes will report stale data for "a
while", where "a while" appears to be easily a couple dozen milliseconds.
This is probably irrelevant in most real-world cases, because the common
case (detecting a node going down) means that we have at least five
seconds after a node goes down before etcd notices the lease expiring,
and a few milliseconds more or less won't matter.
But we have tests that assume either that node 0 is always the
coordinator (wrong) or that waiting for node 0 to think the cluster
is up means that every node in the cluster thinks the cluster is up,
or at least that it means that the coordinator thinks the cluster is
up. We retried later operations but not the initial ones against
the coordinator.
In fact, we probably want to wait for the entire cluster to think
it's up before we start trying things on clusters.
We also replace the "CheckClusterState" function with the existing
AwaitState call, or a new AssertState which errors out since that's
the way we usually use AwaitState anyway.
In the AwaitPrimaryState function, which used to be
AwaitCoordinatorState in a different long-lost revision, we have
to delay until a primary node is available, or fail if one does
not become available, to avoid a panic. This probably shouldn't
happen anymore, because of the last change:
Also, rovide dummy topology.Node entries before metadata is read.
During initial startup, we want to be able to do things like determine
which node is the primary, even before we've read metadata from them.
To do this, we populate the node list with dummy entries that just have
the ID (the only part we need to sort our list), and a node state of
UNKNOWN.
This breaks the fancy logic for determining whether or not to update
the node data, because the initial status of UNKNOWN matches what we
get from SetMetadata giving us new data so we end up not realizing
that this was actually a meaningful change. But actually, that's
a pretty niche optimization; we usually only get state changes when
there's an actual change in state. The updates here are cheap
and only happen after a write (or on the first query) so it's not
worth making the logic a lot fancier to make it work, when we can
just do the simple thing and update any time the dirty flag is set.
We also standardize on a 50ms delay, because 1ms delays were
really expensive when each check was hitting etcd multiple times,
and 50ms is Usually Long Enough.
This is a significant overhaul! Quite a lot of things changed here.
Basically: Prior to this, every request for data from etcd implies
requesting the current live data from etcd, and then unpacking it or
extracting it in some way. This is expensive, which is why we have
a cache in front of it.
We don't need to do that! We can use a Watch, which notifies us
of changes as changes happen. However, there's some challenges and
difficulties along the way, and there's a couple of other changes
which are included here because it's a pain to try to separate them
out.
1. We require a logger to be provided to create our internal Etcd
wrapper. We then use that logger, instead of `fmt.Printf`. This makes
debugging messages work better, and also diagnostics, and so on.
2. The internal client that we are reusing can enter a failed state
after a leader election, in which case we have to recreate the client
to have a working client. We add a new internal-use method,
`retryClient`, which wraps a function which takes an etcd client
and returns an error, and checks for leader-election type errors
and retries creating the client when they happen. That last bit
has not been successfully tested because it's actually really hard
to trigger this now. (Because it was related in part to the
amount of etcd traffic we were producing, which is reduced.)
3. The general swap over from looking things up to unpacking things
as they come in, then returning those already-unpacked things when
we get requests.
With this change, *many tests will fail*. That is addressed by
a separate commit which addresses the secondary problem, which is
that some of our test harness code was relying on the assumption
that if any node in a cluster thinks the cluster is up, every node
will. That was usually true when we were doing everything as
expensive fully-synchronized cluster checks, but becomes significantly
less reliably true in real-world cases where nodes are also
going down sometimes, or nodes are going up and down unexpectedly.
The new etcd implementation has internal caching-like behavior which is
much more reliable (it doesn't use a TTL, it just updates when there's updates
to process) so we don't need this cache.
The existence field wasn't working because runs were broken for filters in
RBF. Fixing that allows us to simplify the logic. Also, we reuse the
findExisting filter because the filter's cached collection of containers
can be reused between things, allowing us to reduce allocations when
there's a lot of views.
The remake container logic (used to avoid allocating extra containers while
applying filters) relied on roaring recomputing N, which it did for bitmaps
but didn't do for runs. Fix this both ways; it would now do that for runs,
but also we add "with explicit N" variants and use those since we have a
correct count already, and don't need it. This means fewer popcounts on
bitmaps, and working at all on runs.
