acl: reduce complexity of token resolution process with alternative singleflighting
Switches acl resolution to use golang.org/x/sync/singleflight. For the
identity/legacy lookups this is a drop-in replacement with the same
overall approach to request coalescing.
For policies this is technically a change in behavior, but when
considered holistically is approximately performance neutral (with the
benefit of less code).
There are two goals with this blob of code (speaking specifically of
policy resolution here):
1) Minimize cross-DC requests.
2) Minimize client-to-server LAN requests.
The previous iteration of this code was optimizing for the case of many
possibly different tokens being resolved concurrently that have a
significant overlap in linked policies such that deduplication would be
worth the complexity. While this is laudable there are some things to
consider that can help to adjust expectations:
1) For v1.4+ policies are always replicated, and once a single policy
shows up in a secondary DC the replicated data is considered
authoritative for requests made in that DC. This means that our
earlier concerns about minimizing cross-DC requests are irrelevant
because there will be no cross-DC policy reads that occur.
2) For Server nodes the in-memory ACL policy cache is capped at zero,
meaning it has no caching. Only Client nodes run with a cache. This
means that instead of having an entire DC's worth of tokens (what a
Server might see) that can have policy resolutions coalesced these
nodes will only ever be seeing node-local token resolutions. In a
reasonable worst-case scenario where a scheduler like Kubernetes has
"filled" a node with Connect services, even that will only schedule
~100 connect services per node. If every service has a unique token
there will only be 100 tokens to coalesce and even then those requests
have to occur concurrently AND be hitting an empty consul cache.
Instead of seeing a great coalescing opportunity for cutting down on
redundant Policy resolutions, in practice it's far more likely given
node densities that you'd see requests for the same token concurrently
than you would for two tokens sharing a policy concurrently (to a degree
that would warrant the overhead of the current variation of
singleflighting.
Given that, this patch switches the Policy resolution process to only
singleflight by requesting token (but keeps the cache as by-policy).
R.B. Boyer