Consider currently dequeued Evaluation's ModifyIndex when determining
its WaitIndex. Normally the Evaluation itself would already be in the
state store snapshot used to determine the WaitIndex. However, since the FSM
applies Raft messages to the state store concurrently with Dequeueing,
it's possible the currently dequeued Evaluation won't yet exist in the
state store snapshot used by JobsForEval.
This can be solved by always considering the current eval's modify index
and using it if it is greater than all of the evals returned by the
state store.
Track the download status of each artifact independently so that if only
one of many artifacts fails to download, completed artifacts aren't
downloaded again.
This seems to fix TestClientAllocations_GarbageCollectAll_Remote being
flaky.
This test confuses me. It joins 2 servers, but then goes out of its way
to make sure the test client only interacts with one. There are not
enough comments for me to figure out the precise assertions this test is
trying to make.
A good old fashioned wait-for-the-client-to-register seems to fix the
flakiness though. The error was that the node could not be found, so
this makes some sense. However, lots of other tests seem to use the same
"wait for node" logic and don't appear to be flaky, so who knows why
waiting fixes this one.
Passes with -race.
Given that the values will rarely change, specially considering that any
changes would be backward incompatible change. As such, it's simpler to
keep syncing manually in the rare occasion and avoid the syncing code
overhead.
nomad/structs is an internal package and imports many libraries (e.g.
raft, codec) that are not relevant to api clients, and may cause
unnecessary dependency pain (e.g. `github.com/ugorji/go/codec`
version is very old now).
Here, we add a code generator that imports the relevant constants from
`nomad/structs`.
I considered using this approach for other structs, but didn't find a
quick viable way to reduce duplication. `nomad/structs` use values as
struct fields (e.g. `string`), while `api` uses value pointer (e.g.
`*string`) instead. Also, sometimes, `api` structs contain deprecated
fields or additional documentation, so simple copy-paste doesn't work.
For these reasons, I opt to keep the status quo.
