This continues the work done in #14908 where a crude solution to prevent a
goroutine leak was implemented. The former code would launch a perpetual
goroutine family every iteration (+1 +1) and the fixed code simply caused a
new goroutine family to first cancel the prior one to prevent the
leak (-1 +1 == 0).
This PR refactors this code completely to:
- make it more understandable
- remove the recursion-via-goroutine strangeness
- prevent unnecessary RPC fetches when the prior one has errored.
The core issue arose from a conflation of the entry.Fetching field to mean:
- there is an RPC (blocking query) in flight right now
- there is a goroutine running to manage the RPC fetch retry loop
The problem is that the goroutine-leak-avoidance check would treat
Fetching like (2), but within the body of a goroutine it would flip that
boolean back to false before the retry sleep. This would cause a new
chain of goroutines to launch which #14908 would correct crudely.
The refactored code uses a plain for-loop and changes the semantics
to track state for "is there a goroutine associated with this cache entry"
instead of the former.
We use a uint64 unique identity per goroutine instead of a boolean so
that any orphaned goroutines can tell when they've been replaced when
the expiry loop deletes a cache entry while the goroutine is still running
and is later replaced.
There is a bug in the error handling code for the Agent cache subsystem discovered:
1. NotifyCallback calls notifyBlockingQuery which calls getWithIndex in
a loop (which backs off on-error up to 1 minute)
2. getWithIndex calls fetch if there’s no valid entry in the cache
3. fetch starts a goroutine which calls Fetch on the cache-type, waits
for a while (again with backoff up to 1 minute for errors) and then
calls fetch to trigger a refresh
The end result being that every 1 minute notifyBlockingQuery spawns an
ancestry of goroutines that essentially lives forever.
This PR ensures that the goroutine started by `fetch` cancels any prior
goroutine spawned by the same line for the same key.
In isolated testing where a cache type was tweaked to indefinitely
error, this patch prevented goroutine counts from skyrocketing.
set -euo pipefail
unset CDPATH
cd "$(dirname "$0")"
for f in $(git grep '\brequire := require\.New(' | cut -d':' -f1 | sort -u); do
echo "=== require: $f ==="
sed -i '/require := require.New(t)/d' $f
# require.XXX(blah) but not require.XXX(tblah) or require.XXX(rblah)
sed -i 's/\brequire\.\([a-zA-Z0-9_]*\)(\([^tr]\)/require.\1(t,\2/g' $f
# require.XXX(tblah) but not require.XXX(t, blah)
sed -i 's/\brequire\.\([a-zA-Z0-9_]*\)(\(t[^,]\)/require.\1(t,\2/g' $f
# require.XXX(rblah) but not require.XXX(r, blah)
sed -i 's/\brequire\.\([a-zA-Z0-9_]*\)(\(r[^,]\)/require.\1(t,\2/g' $f
gofmt -s -w $f
done
for f in $(git grep '\bassert := assert\.New(' | cut -d':' -f1 | sort -u); do
echo "=== assert: $f ==="
sed -i '/assert := assert.New(t)/d' $f
# assert.XXX(blah) but not assert.XXX(tblah) or assert.XXX(rblah)
sed -i 's/\bassert\.\([a-zA-Z0-9_]*\)(\([^tr]\)/assert.\1(t,\2/g' $f
# assert.XXX(tblah) but not assert.XXX(t, blah)
sed -i 's/\bassert\.\([a-zA-Z0-9_]*\)(\(t[^,]\)/assert.\1(t,\2/g' $f
# assert.XXX(rblah) but not assert.XXX(r, blah)
sed -i 's/\bassert\.\([a-zA-Z0-9_]*\)(\(r[^,]\)/assert.\1(t,\2/g' $f
gofmt -s -w $f
done
* Fix bug in cache where TTLs are effectively ignored
This mostly affects streaming since streaming will immediately return from Fetch calls when the state is Closed on eviction which causes the race condition every time.
However this also affects all other cache types if the fetch call happens to return between the eviction and then next time around the Get loop by any client.
There is a separate bug that allows cache items to be evicted even when there are active clients which is the trigger here.
* Add changelog entry
* Update .changelog/9978.txt
Add a skip condition to all tests slower than 100ms.
This change was made using `gotestsum tool slowest` with data from the
last 3 CI runs of master.
See https://github.com/gotestyourself/gotestsum#finding-and-skipping-slow-tests
With this change:
```
$ time go test -count=1 -short ./agent
ok github.com/hashicorp/consul/agent 0.743s
real 0m4.791s
$ time go test -count=1 -short ./agent/consul
ok github.com/hashicorp/consul/agent/consul 4.229s
real 0m8.769s
```
Prepopulate was setting entry.Expiry.HeapIndex to 0. Previously this would result in a call to heap.Fix(0)
which wasn't correct, but was also not really a problem because at worse it would re-notify.
With the recent change to extract cachettl it was changed to call Update(idx), which would have updated
the wrong entry.
A previous commit removed the setting of entry.Expiry so that the HeapIndex would be reported
as -1, and this commit adds a test and handles the -1 heap index.
The rationale behind removing them is that all of our own code (xDS, builtin connect proxy) use the cache notification mechanism. This ensures that the blocking fetch behind the scenes is always executing. Therefore the only way you might go to get a certificate and have to wait is when 1) the request has never been made for that cert before or 2) you are using the v1/agent/connect/ca/leaf API for retrieving the cert yourself.
In the first case, the refresh change doesn’t alter the behavior. In the second case, it can be mitigated by using blocking queries with that API which just like normal cache notification mechanism will cause the blocking fetch to be initiated and to get leaf certs as soon as needed.
If you are not using blocking queries, or Envoy/xDS, or the builtin connect proxy but are retrieving the certs yourself then the HTTP endpoint might take a little longer to respond.
This also renames the RefreshTimeout field on the register options to QueryTimeout to more accurately reflect that it is used for any type that supports blocking queries.
Blocking queries issues will still be uncancellable (that cannot be helped until we get rid of net/rpc). However this makes it so that if calling getWithIndex (like during a cache Notify go routine) we can cancell the outer routine. Previously it would keep issuing more blocking queries until the result state actually changed.
Previously the SupportsBlocking option was specified by a method on the
type, and all the other options were specified from RegisterOptions.
This change moves RegisterOptions to a method on the type, and moves
SupportsBlocking into the options struct.
Currently there are only 2 cache-types. So all cache-types can implement
this method by embedding a struct with those predefined values. In the
future if a cache type needs to be registered more than once with different
options it can remove the embedded type and implement the method in a way
that allows for paramaterization.
* Fix race condition during a cache get
Check the entry we pulled out of the cache while holding the lock had Fetching set.
If it did then we should use the existing Waiter instead of calling fetch. The reason
this is better than just calling fetch is that fetch re-gets the entry out of the
entries map and the previous fetch may have finished. Therefore this prevents
erroneously starting a new fetch because we just missed the last update.
* Fix race condition fully
The first commit still allowed for the following scenario:
• No entry existing when checked in getWithIndex while holding the read lock
• Then by time we had reached fetch it had been created and finished.
* always use ok when returning
* comment mentioning the reading from entries.
* use cacheHit consistently
* Support rate limiting and concurrency limiting CSR requests on servers; handle CA rotations gracefully with jitter and backoff-on-rate-limit in client
* Add CSR rate limiting docs
* Fix config naming and add tests for new CA configs
* Add State storage and LastResult argument into Cache so that cache.Types can safely store additional data that is eventually expired.
* New Leaf cache type working and basic tests passing. TODO: more extensive testing for the Root change jitter across blocking requests, test concurrent fetches for different leaves interact nicely with rootsWatcher.
* Add multi-client and delayed rotation tests.
* Typos and cleanup error handling in roots watch
* Add comment about how the FetchResult can be used and change ca leaf state to use a non-pointer state.
* Plumb test override of root CA jitter through TestAgent so that tests are deterministic again!
* Fix failing config test
* Fixes#4480. Don't leave old errors around in cache that can be hit in specific circumstances.
* Move error reset to cover extreme edge case of nil Value, nil err Fetch
* Add cache types for catalog/services and health/services and basic test that caching works
* Support non-blocking cache types with Cache-Control semantics.
* Update API docs to include caching info for every endpoint.
* Comment updates per PR feedback.
* Add note on caching to the 10,000 foot view on the architecture page to make the new data path more clear.
* Document prepared query staleness quirk and force all background requests to AllowStale so we can spread service discovery load across servers.
* Fix theoretical cache collision bug if/when we use more cache types with same result type
* Generalized fix for blocking query handling when state store methods return zero index
* Refactor test retry to only affect CI
* Undo make file merge
* Add hint to error message returned to end-user requests if Connect is not enabled when they try to request cert
* Explicit error for Roots endpoint if connect is disabled
* Fix tests that were asserting old behaviour
Few other fixes in here just to get a clean run locally - they are all also fixed in other PRs but shouldn't conflict.
This should be robust to timing between goroutines now.
Ronald
R.B. Boyer
R.B. Boyer
Paul Banks
Matt Keeler
Daniel Nephin
Pierre Souchay
Hans Hasselberg
Matt Keeler
Mitchell Hashimoto