Files
moloapp/src/supervisor/TOPOLOGY.md
T

99 lines
3.9 KiB
Markdown
Raw Normal View History

# Multi-instance topology (S7-9, E7-8)
N stateless app instances behind a load balancer, one shared durable bus,
one shared session and KV store, and one scheduler leader lease in shared
KV. Anything needed by more than one instance lives in a shared backend;
local memory is never treated as global (R7-8).
## Pieces owned by this part
- `molo::bus::TopicBus`: ephemeral broadcast fan-out (at-most-once).
- `molo::bus::StreamBus`: durable streams for catch-up replay (at-least-once).
- `molo::tasks::Scheduler`: single leader via the shared-KV lease, passive
followers, last-fire records in shared KV, idempotency-keyed fires.
- `molo::replication::Replicator` plus `SyncSession`: oplog push plus
catch-up pull between instances with per-model conflict policies.
- `molo::supervisor::Supervisor::status_json()`: health input for endpoints.
## Pieces owned elsewhere
- 04-data.md: the shared session store and shared KV backend. This part
programs against `molo::common::KvStore`; `InMemoryKvStore` is the
single-node implementation. Multi-instance deployments swap in the
shared backend without touching scheduler or queue code.
- 06-realtime.md: room and broadcast semantics. The rule is: broadcast
emit publishes once to the bus; every node re-emits to its local
sockets. See `tests/src/supervisor/test_multi_instance.cpp`.
- 09-cli-ops.md and 05-request-routing.md: health/readiness endpoints and
the CLI surface. They read `status_json()` and drive `Supervisor` plus
`TaskQueue` only through the declared APIs (R7-10).
## Load balancer: sticky sessions
WebSocket and Socket.IO long-polling transports must pin to one node (BP-6,
S7-9), otherwise frames for one connection land on different nodes. Plain
WS and SSE may pin or re-resolve per route (documented in 06-realtime.md).
Example (haproxy, cookie-based stickiness):
```
backend molo_nodes
balance roundrobin
cookie MOLO_NODE insert indirect nocache
server node1 10.0.0.11:8080 cookie n1 check
server node2 10.0.0.12:8080 cookie n2 check
```
Example (nginx, ip_hash pinning):
```
upstream molo_nodes {
ip_hash;
server 10.0.0.11:8080;
server 10.0.0.12:8080;
}
```
Health checks should hit the readiness endpoint (wired by 09-cli-ops.md
from `status_json()`, queue depth, and sync lag gauges) and shed a node
whose supervisor reports fatal workers or whose lag passes 10000 (S7-7).
## Broadcast and catch-up pattern
```cpp
// Emit once (any node):
live.publish("room:lobby", payload); // ephemeral fan-out
bus.publish("roomlog", payload); // durable copy for reconnects
// Every live node re-emits to its local sockets:
for (auto& m : live.poll(sub, 64)) local_emit(m.payload);
// A restarted node catches up what it missed while dead:
for (auto& m : bus.poll("roomlog", node_id, 64)) { replay(m); ack(m.seq); }
```
## Scheduler leadership
One leader per deployment (S7-5): lease key `molo/scheduler/leader`,
10s TTL, 3s renewal. Follower takeover is bounded by the TTL (proven in
`test_scheduler.cpp`). Dual-leader overlap during partitions is accepted
with bounds (O7-4): overlapping fires carry the same idempotency key, so
the queue dedupes them and no duplicate run happens.
## Known v1 limits
- Thread workers only; OS-process isolation is a v2 transport behind the
same `WorkFunc` interface (O7-2).
- In-house backends only; no Redis-protocol compatibility (O7-1).
- At-least-once everywhere; no exactly-once primitive (O7-3). Handlers
must be idempotent (R7-2).
- The dead-letter queue is in-memory per process; entries do not survive
a process restart yet. Queued (non-dead) work does survive via the
durable op log.
- The oplog keeps full history per instance; operators archive or trim
`oplog.log` files on a schedule. Bounded per-round sync batches keep
memory flat regardless of log size.
- Durable streams use a dedicated append-only segment file per stream.
Sharing the 04-data.md SQLite WAL path instead is still an open joint
decision (O7-5); the `SegmentLog` seam keeps the swap local.