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.
This includes fixes that improve gossip scalability on very large (> 10k node) clusters.
The Serf changes:
- take snapshot disk IO out of the critical path for handling messages hashicorp/serf#524
- make snapshot compaction much less aggressive - the old fixed threshold caused snapshots to be constantly compacted (synchronously with request handling) on clusters larger than about 2000 nodes! hashicorp/serf#525
Memberlist changes:
- prioritize handling alive messages over suspect/dead to improve stability, and handle queue in LIFO order to avoid acting on info that 's already stale in the queue by the time we handle it. hashicorp/memberlist#159
- limit the number of concurrent pushPull requests being handled at once to 128. In one test scenario with 10s of thousands of servers we saw channel and lock blocking cause over 3000 pushPulls at once which ballooned the memory of the server because each push pull contained a de-serialised list of all known 10k+ nodes and their tags for a total of about 60 million objects and 7GB of memory stuck. While the rest of the fixes here should prevent the same root cause from blocking in the same way, this prevents any other bug or source of contention from allowing pushPull messages to stack up and eat resources. hashicorp/memberlist#158
This gets a number of improvements:
* Fixed a missing case where gossip would stop flowing to dead nodes for a short while.
* Uses the go-sockaddr library to look for private IP addresses, which prefers non-loopback
private addresses over loopback ones when trying to automatically determine the advertise address.
* Properly seeds Go's random number generator using the seed library.
* Serf snapshots no longer have the executable bit set on thie file.
R.B. Boyer