Registering gRPC balancers is thread-unsafe because they are stored in a
global map variable that is accessed without holding a lock. Therefore,
it's expected that balancers are registered _once_ at the beginning of
your program (e.g. in a package `init` function) and certainly not after
you've started dialing connections, etc.
> NOTE: this function must only be called during initialization time
> (i.e. in an init() function), and is not thread-safe.
While this is fine for us in production, it's challenging for tests that
spin up multiple agents in-memory. We currently register a balancer per-
agent which holds agent-specific state that cannot safely be shared.
This commit introduces our own registry that _is_ thread-safe, and
implements the Builder interface such that we can call gRPC's `Register`
method once, on start-up. It uses the same pattern as our resolver
registry where we use the dial target's host (aka "authority"), which is
unique per-agent, to determine which builder to use.
Protobuf Refactoring for Multi-Module Cleanliness
This commit includes the following:
Moves all packages that were within proto/ to proto/private
Rewrites imports to account for the packages being moved
Adds in buf.work.yaml to enable buf workspaces
Names the proto-public buf module so that we can override the Go package imports within proto/buf.yaml
Bumps the buf version dependency to 1.14.0 (I was trying out the version to see if it would get around an issue - it didn't but it also doesn't break things and it seemed best to keep up with the toolchain changes)
Why:
In the future we will need to consume other protobuf dependencies such as the Google HTTP annotations for openapi generation or grpc-gateway usage.
There were some recent changes to have our own ratelimiting annotations.
The two combined were not working when I was trying to use them together (attempting to rebase another branch)
Buf workspaces should be the solution to the problem
Buf workspaces means that each module will have generated Go code that embeds proto file names relative to the proto dir and not the top level repo root.
This resulted in proto file name conflicts in the Go global protobuf type registry.
The solution to that was to add in a private/ directory into the path within the proto/ directory.
That then required rewriting all the imports.
Is this safe?
AFAICT yes
The gRPC wire protocol doesn't seem to care about the proto file names (although the Go grpc code does tack on the proto file name as Metadata in the ServiceDesc)
Other than imports, there were no changes to any generated code as a result of this.
* feat: calculate retry wait time with exponential back-off
* test: add test for getWaitTime method
* feat: enforce random jitter between min value from previous iteration and current
* extract randomStagger to simplify tests and use Milliseconds to avoid float math.
* rename variables
* add test and rename comment
---------
Co-authored-by: Poonam Jadhav <poonam.jadhav@hashicorp.com>
- Fixes a panic when Operation.SourceAddr is nil (internal net/rpc calls)
- Adds proper HTTP response codes (429 and 503) for rate limit errors
- Makes the error messages clearer
- Enables automatic retries for rate-limit errors in the net/rpc stack
This is the OSS portion of enterprise PR 3822.
Adds a custom gRPC balancer that replicates the router's server cycling
behavior. Also enables automatic retries for RESOURCE_EXHAUSTED errors,
which we now get for free.
* update go version to 1.18 for api and sdk, go mod tidy
* removes ioutil usage everywhere which was deprecated in go1.16 in favour of io and os packages. Also introduces a lint rule which forbids use of ioutil going forward.
Co-authored-by: R.B. Boyer <4903+rboyer@users.noreply.github.com>
Fix an issue where rpc_hold_timeout was being used as the timeout for non-blocking queries. Users should be able to tune read timeouts without fiddling with rpc_hold_timeout. A new configuration `rpc_read_timeout` is created.
Refactor some implementation from the original PR 11500 to remove the misleading linkage between RPCInfo's timeout (used to retry in case of certain modes of failures) and the client RPC timeouts.
Previously, public referred to gRPC services that are both exposed on
the dedicated gRPC port and have their definitions in the proto-public
directory (so were considered usable by 3rd parties). Whereas private
referred to services on the multiplexed server port that are only usable
by agents and other servers.
Now, we're splitting these definitions, such that external/internal
refers to the port and public/private refers to whether they can be used
by 3rd parties.
This is necessary because the peering replication API needs to be
exposed on the dedicated port, but is not (yet) suitable for use by 3rd
parties.
This is the OSS portion of enterprise PR 2056.
This commit provides server-local implementations of the proxycfg.ConfigEntry
and proxycfg.ConfigEntryList interfaces, that source data from streaming events.
It makes use of the LocalMaterializer type introduced for peering replication,
adding the necessary support for authorization.
It also adds support for "wildcard" subscriptions (within a topic) to the event
publisher, as this is needed to fetch service-resolvers for all services when
configuring mesh gateways.
Currently, events will be emitted for just the ingress-gateway, service-resolver,
and mesh config entry types, as these are the only entries required by proxycfg
— the events will be emitted on topics named IngressGateway, ServiceResolver,
and MeshConfig topics respectively.
Though these events will only be consumed "locally" for now, they can also be
consumed via the gRPC endpoint (confirmed using grpcurl) so using them from
client agents should be a case of swapping the LocalMaterializer for an
RPCMaterializer.
Adds a timeout (deadline) to client RPC calls, so that streams will no longer hang indefinitely in unstable network conditions.
Co-authored-by: kisunji <ckim@hashicorp.com>
* Fixes a lint warning about t.Errorf not supporting %w
* Enable running autopilot on all servers
On the non-leader servers all they do is update the state and do not attempt any modifications.
* Fix the RPC conn limiting tests
Technically they were relying on racey behavior before. Now they should be reliable.
Introduces the capability to configure TLS differently for Consul's
listeners/ports (i.e. HTTPS, gRPC, and the internal multiplexed RPC
port) which is useful in scenarios where you may want the HTTPS or
gRPC interfaces to present a certificate signed by a well-known/public
CA, rather than the certificate used for internal communication which
must have a SAN in the form `server.<dc>.consul`.
Improves tests from #12362
These tests try to setup the following concurrent scenario:
1. (goroutine 1) execute read RPC with index=0
2. (goroutine 1) get response from (1) @ index=10
3. (goroutine 1) execute read RPC with index=10 and block
4. (goroutine 2) WHILE (3) is blocking, start slamming the system with stray writes that will cause the WatchSet to wakeup
5. (goroutine 2) after doing all writes, shut down the reader above
6. (goroutine 1) stops reading, double checks that it only ever woke up once (from 1)
Starting from and extending the mechanism introduced in #12110 we can specially handle the 3 main special Consul RPC endpoints that react to many config entries in a single blocking query in Connect:
- `DiscoveryChain.Get`
- `ConfigEntry.ResolveServiceConfig`
- `Intentions.Match`
All of these will internally watch for many config entries, and at least one of those will likely be not found in any given query. Because these are blends of multiple reads the exact solution from #12110 isn't perfectly aligned, but we can tweak the approach slightly and regain the utility of that mechanism.
### No Config Entries Found
In this case, despite looking for many config entries none may be found at all. Unlike #12110 in this scenario we do not return an empty reply to the caller, but instead synthesize a struct from default values to return. This can be handled nearly identically to #12110 with the first 1-2 replies being non-empty payloads followed by the standard spurious wakeup suppression mechanism from #12110.
### No Change Since Last Wakeup
Once a blocking query loop on the server has completed and slept at least once, there is a further optimization we can make here to detect if any of the config entries that were present at specific versions for the prior execution of the loop are identical for the loop we just woke up for. In that scenario we can return a slightly different internal sentinel error and basically externally handle it similar to #12110.
This would mean that even if 20 discovery chain read RPC handling goroutines wakeup due to the creation of an unrelated config entry, the only ones that will terminate and reply with a blob of data are those that genuinely have new data to report.
### Extra Endpoints
Since this pattern is pretty reusable, other key config-entry-adjacent endpoints used by `agent/proxycfg` also were updated:
- `ConfigEntry.List`
- `Internal.IntentionUpstreams` (tproxy)
By using the query results as state.
Blocking queries are efficient when the query matches some results,
because the ModifyIndex of those results, returned as queryMeta.Mindex,
will never change unless the items themselves change.
Blocking queries for non-existent items are not efficient because the
queryMeta.Index can (and often does) change when other entities are
written.
This commit reduces the churn of these queries by using a different
comparison for "has changed". Instead of using the modified index, we
use the existence of the results. If the previous result was "not found"
and the new result is still "not found", we know we can ignore the
modified index and continue to block.
This is done by setting the minQueryIndex to the returned
queryMeta.Index, which prevents the query from returning before a state
change is observed.
This commit syncs ENT changes to the OSS repo.
Original commit details in ENT:
```
commit 569d25f7f4578981c3801e6e067295668210f748
Author: FFMMM <FFMMM@users.noreply.github.com>
Date: Thu Feb 10 10:23:33 2022 -0800
Vendor fork net rpc (#1538)
* replace net/rpc w consul-net-rpc/net/rpc
Signed-off-by: FFMMM <FFMMM@users.noreply.github.com>
* replace msgpackrpc and go-msgpack with fork from mono repo
Signed-off-by: FFMMM <FFMMM@users.noreply.github.com>
* gofmt all files touched
Signed-off-by: FFMMM <FFMMM@users.noreply.github.com>
```
Signed-off-by: FFMMM <FFMMM@users.noreply.github.com>
This field was documented as enabling TLS for outgoing RPC, but that was not the case.
All this field did was set the use_tls serf tag.
Instead of setting this field in a place far from where it is used, move the logic to where
the serf tag is set, so that the code is much more obvious.
tlsutil.Config already presents an excellent structure for this
configuration. Copying the runtime config fields to agent/consul.Config
makes code harder to trace, and provides no advantage.
Instead of copying the fields around, use the tlsutil.Config struct
directly instead.
This is one small step in removing the many layers of duplicate
configuration.
Also accept an RPCInfo instead of interface{}. Accepting an interface
lead to a bug where the caller was expecting the arg to be the response
when in fact it was always passed the request. By accepting RPCInfo
it should indicate that this is actually the request value.
One caller of canRetry already passed an RPCInfo, the second handles
the type assertion before calling canRetry.
I believe this commit also fixes a bug. Previously RPCMaxConnsPerClient was not being re-read from the RuntimeConfig, so passing it to Server.ReloadConfig was never changing the value.
Also improve the test runtime by not doing a lot of unnecessary work.
This can happen when one other node in the cluster such as a client is unable to communicate with the leader server and sees it as failed. When that happens its failing status eventually gets propagated to the other servers in the cluster and eventually this can result in RPCs returning “No cluster leader” error.
That error is misleading and unhelpful for determing the root cause of the issue as its not raft stability but rather and client -> server networking issue. Therefore this commit will add a new error that will be returned in that case to differentiate between the two cases.
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
```
This is like a Möbius strip of code due to the fact that low-level components (serf/memberlist) are connected to high-level components (the catalog and mesh-gateways) in a twisty maze of references which make it hard to dive into. With that in mind here's a high level summary of what you'll find in the patch:
There are several distinct chunks of code that are affected:
* new flags and config options for the server
* retry join WAN is slightly different
* retry join code is shared to discover primary mesh gateways from secondary datacenters
* because retry join logic runs in the *agent* and the results of that
operation for primary mesh gateways are needed in the *server* there are
some methods like `RefreshPrimaryGatewayFallbackAddresses` that must occur
at multiple layers of abstraction just to pass the data down to the right
layer.
* new cache type `FederationStateListMeshGatewaysName` for use in `proxycfg/xds` layers
* the function signature for RPC dialing picked up a new required field (the
node name of the destination)
* several new RPCs for manipulating a FederationState object:
`FederationState:{Apply,Get,List,ListMeshGateways}`
* 3 read-only internal APIs for debugging use to invoke those RPCs from curl
* raft and fsm changes to persist these FederationStates
* replication for FederationStates as they are canonically stored in the
Primary and replicated to the Secondaries.
* a special derivative of anti-entropy that runs in secondaries to snapshot
their local mesh gateway `CheckServiceNodes` and sync them into their upstream
FederationState in the primary (this works in conjunction with the
replication to distribute addresses for all mesh gateways in all DCs to all
other DCs)
* a "gateway locator" convenience object to make use of this data to choose
the addresses of gateways to use for any given RPC or gossip operation to a
remote DC. This gets data from the "retry join" logic in the agent and also
directly calls into the FSM.
* RPC (`:8300`) on the server sniffs the first byte of a new connection to
determine if it's actually doing native TLS. If so it checks the ALPN header
for protocol determination (just like how the existing system uses the
type-byte marker).
* 2 new kinds of protocols are exclusively decoded via this native TLS
mechanism: one for ferrying "packet" operations (udp-like) from the gossip
layer and one for "stream" operations (tcp-like). The packet operations
re-use sockets (using length-prefixing) to cut down on TLS re-negotiation
overhead.
* the server instances specially wrap the `memberlist.NetTransport` when running
with gateway federation enabled (in a `wanfed.Transport`). The general gist is
that if it tries to dial a node in the SAME datacenter (deduced by looking
at the suffix of the node name) there is no change. If dialing a DIFFERENT
datacenter it is wrapped up in a TLS+ALPN blob and sent through some mesh
gateways to eventually end up in a server's :8300 port.
* a new flag when launching a mesh gateway via `consul connect envoy` to
indicate that the servers are to be exposed. This sets a special service
meta when registering the gateway into the catalog.
* `proxycfg/xds` notice this metadata blob to activate additional watches for
the FederationState objects as well as the location of all of the consul
servers in that datacenter.
* `xds:` if the extra metadata is in place additional clusters are defined in a
DC to bulk sink all traffic to another DC's gateways. For the current
datacenter we listen on a wildcard name (`server.<dc>.consul`) that load
balances all servers as well as one mini-cluster per node
(`<node>.server.<dc>.consul`)
* the `consul tls cert create` command got a new flag (`-node`) to help create
an additional SAN in certs that can be used with this flavor of federation.
