There are some cross-config-entry relationships that are enforced during
"graph validation" at persistence time that are required to be
maintained. This means that config entries may form a digraph at times.
Config entry replication procedes in a particular sorted order by kind
and name.
Occasionally there are some fixups to these digraphs that end up
replicating in the wrong order and replicating the leaves
(ingress-gateway) before the roots (service-defaults) leading to
replication halting due to a graph validation error related to things
like mismatched service protocol requirements.
This PR changes replication to give each computed change (upsert/delete)
a fair shot at being applied before deciding to terminate that round of
replication in error. In the case where we've simply tried to do the
operations in the wrong order at least ONE of the outstanding requests
will complete in the right order, leading the subsequent round to have
fewer operations to do, with a smaller likelihood of graph validation
errors.
This does not address all scenarios, but for scenarios where the edits
are being applied in the wrong order this should avoid replication
halting.
Fixes#9319
The scenario that is NOT ADDRESSED by this PR is as follows:
1. create: service-defaults: name=new-web, protocol=http
2. create: service-defaults: name=old-web, protocol=http
3. create: service-resolver: name=old-web, redirect-to=new-web
4. delete: service-resolver: name=old-web
5. update: service-defaults: name=old-web, protocol=grpc
6. update: service-defaults: name=new-web, protocol=grpc
7. create: service-resolver: name=old-web, redirect-to=new-web
If you shutdown dc2 just before (4) and turn it back on after (7)
replication is impossible as there is no single edit you can make to
make forward progress.
Previously we were inconsistently checking the response for errors. This
PR moves the response-is-error check into raftApply, so that all callers
can look at only the error response, instead of having to know that
errors could come from two places.
This should expose a few more errors that were previously hidden because
in some calls to raftApply we were ignoring the response return value.
Also handle errors more consistently. In some cases we would log the
error before returning it. This can be very confusing because it can
result in the same error being logged multiple times. Instead return
a wrapped error.
Previously config entries sharing a kind & name but in different
namespaces could occasionally cause "stuck states" in replication
because the namespace fields were ignored during the differential
comparison phase.
Example:
Two config entries written to the primary:
kind=A,name=web,namespace=bar
kind=A,name=web,namespace=foo
Under the covers these both get saved to memdb, so they are sorted by
all 3 components (kind,name,namespace) during natural iteration. This
means that before the replication code does it's own incomplete sort,
the underlying data IS sorted by namespace ascending (bar comes before
foo).
After one pass of replication the primary and secondary datacenters have
the same set of config entries present. If
"kind=A,name=web,namespace=bar" were to be deleted, then things get
weird. Before replication the two sides look like:
primary: [
kind=A,name=web,namespace=foo
]
secondary: [
kind=A,name=web,namespace=bar
kind=A,name=web,namespace=foo
]
The differential comparison phase walks these two lists in sorted order
and first compares "kind=A,name=web,namespace=foo" vs
"kind=A,name=web,namespace=bar" and falsely determines they are the SAME
and are thus cause an update of "kind=A,name=web,namespace=foo". Then it
compares "<nothing>" with "kind=A,name=web,namespace=foo" and falsely
determines that the latter should be DELETED.
During reconciliation the deletes are processed before updates, and so
for a brief moment in the secondary "kind=A,name=web,namespace=foo" is
erroneously deleted and then immediately restored.
Unfortunately after this replication phase the final state is identical
to the initial state, so when it loops around again (rate limited) it
repeats the same set of operations indefinitely.
- Upgrade the ConfigEntry.ListAll RPC to be kind-aware so that older
copies of consul will not see new config entries it doesn't understand
replicate down.
- Add shim conversion code so that the old API/CLI method of interacting
with intentions will continue to work so long as none of these are
edited via config entry endpoints. Almost all of the read-only APIs will
continue to function indefinitely.
- Add new APIs that operate on individual intentions without IDs so that
the UI doesn't need to implement CAS operations.
- Add a new serf feature flag indicating support for
intentions-as-config-entries.
- The old line-item intentions way of interacting with the state store
will transparently flip between the legacy memdb table and the config
entry representations so that readers will never see a hiccup during
migration where the results are incomplete. It uses a piece of system
metadata to control the flip.
- The primary datacenter will begin migrating intentions into config
entries on startup once all servers in the datacenter are on a version
of Consul with the intentions-as-config-entries feature flag. When it is
complete the old state store representations will be cleared. We also
record a piece of system metadata indicating this has occurred. We use
this metadata to skip ALL of this code the next time the leader starts
up.
- The secondary datacenters continue to run the old intentions
replicator until all servers in the secondary DC and primary DC support
intentions-as-config-entries (via serf flag). Once this condition it met
the old intentions replicator ceases.
- The secondary datacenters replicate the new config entries as they are
migrated in the primary. When they detect that the primary has zeroed
it's old state store table it waits until all config entries up to that
point are replicated and then zeroes its own copy of the old state store
table. We also record a piece of system metadata indicating this has
occurred. We use this metadata to skip ALL of this code the next time
the leader starts up.
