Files
multica/server/internal/daemon/wsrpc.go
ZIce 92676fbe51 fix(daemon): recover from uncertain ws claims
Squash merge approved after review and regression verification.
2026-07-20 11:50:06 +08:00

381 lines
14 KiB
Go

package daemon
import (
"context"
"encoding/json"
"errors"
"fmt"
"sync"
"time"
"github.com/google/uuid"
"github.com/multica-ai/multica/server/pkg/protocol"
)
// errWSRPCUnavailable is returned by wsRPCClient.Call when there is no live WS
// connection to carry the request. Callers treat it as the signal to fall back
// to HTTP.
var errWSRPCUnavailable = errors.New("ws rpc: no active connection")
// errWSRPCUncertain is returned when a request's frame WAS sent but the
// connection dropped before a definitive response. The outcome is unknown (the
// server may have committed), so the caller must NOT fall back to another
// transport for the same work — that risks a double claim (MUL-4257).
var errWSRPCUncertain = errors.New("ws rpc: sent but outcome unknown (connection lost)")
// wsRPCResponseGrace is how much longer the daemon waits for an RPC response
// beyond the server-side execution budget it requested, so a claim that
// committed just before the server deadline still reports back before the
// daemon gives up (MUL-4257).
const wsRPCResponseGrace = 2 * time.Second
var wsClaimUncertainFallbackDelay = batchClaimRequestTimeout + wsRPCResponseGrace
// errWSRPCWriteBufferFull is returned when the connection's write buffer is
// saturated; the caller falls back to HTTP rather than blocking the socket.
var errWSRPCWriteBufferFull = errors.New("ws rpc: write buffer full")
// wsRPCClient is the daemon-side half of the generic WS request/response
// transport (MUL-4257). It correlates responses to requests by request_id over
// the shared, multiplexed WS control connection so multiple RPCs can be in
// flight concurrently. Sending is delegated to an injected sendFrame func
// (which pushes onto the active connection's write channel); when no connection
// is attached, Call fails fast with errWSRPCUnavailable and the caller uses
// HTTP.
// wsOutbound is a frame queued for the WS writer. It is cancelable so an RPC
// caller that gives up (timeout/detach) before the frame has hit the socket can
// prevent it from being delivered later — otherwise a backpressured writer
// could deliver a stale tasks.claim after the daemon already HTTP-fell-back,
// double-claiming (MUL-4257, Sol-Boy review). sent/cancel race under mu so the
// decision is atomic: whoever wins determines whether the frame is delivered.
type wsOutbound struct {
data []byte
mu sync.Mutex
sent bool
canceled bool
}
// beginWrite is called by the writer immediately before WriteMessage. It
// returns false when the frame was already cancelled (skip it); otherwise it
// marks the frame sent so a concurrent cancel() can no longer un-send it.
func (o *wsOutbound) beginWrite() bool {
o.mu.Lock()
defer o.mu.Unlock()
if o.canceled {
return false
}
o.sent = true
return true
}
// cancel is called by an RPC caller giving up. Returns true if the frame was
// still pending (now cancelled — the writer will skip it, so it is guaranteed
// NOT delivered); false if the writer already began sending it.
func (o *wsOutbound) cancel() bool {
o.mu.Lock()
defer o.mu.Unlock()
if o.sent {
return false
}
o.canceled = true
return true
}
type wsRPCClient struct {
mu sync.Mutex
pending map[string]chan protocol.RPCResponsePayload
sendFrame func([]byte) (*wsOutbound, error)
// rpcV1Supported belongs to the currently attached connection. attach
// clears it before exposing a replacement sender so a claim that races a
// reconnect can never carry negotiation state across connections.
rpcV1Supported bool
generation uint64
// grace is added to a call's server-side timeout budget to get how long the
// daemon waits for the response, so a claim that committed just before the
// server deadline still reports back before the daemon gives up (MUL-4257).
grace time.Duration
}
func newWSRPCClient(grace time.Duration) *wsRPCClient {
return &wsRPCClient{
pending: make(map[string]chan protocol.RPCResponsePayload),
grace: grace,
}
}
// attach binds a live connection's frame writer and clears the previous
// connection's negotiated capability. Passing nil detaches (on disconnect),
// after which Call fails fast until the next attach and rpc-v1 heartbeat ack.
// Any pending requests are failed so their callers fall back to HTTP
// immediately.
func (c *wsRPCClient) attach(sendFrame func([]byte) (*wsOutbound, error)) uint64 {
c.mu.Lock()
c.generation++
c.rpcV1Supported = false
c.sendFrame = sendFrame
for id, ch := range c.pending {
close(ch)
delete(c.pending, id)
}
generation := c.generation
c.mu.Unlock()
return generation
}
// markRPCV1Supported records explicit server support for the currently
// attached connection. Heartbeat acks received without a live sender cannot
// enable a future connection.
func (c *wsRPCClient) markRPCV1Supported(generation uint64) {
if c == nil {
return
}
c.mu.Lock()
if c.sendFrame != nil && c.generation == generation {
c.rpcV1Supported = true
}
c.mu.Unlock()
}
func (c *wsRPCClient) currentGeneration() uint64 {
if c == nil {
return 0
}
c.mu.Lock()
defer c.mu.Unlock()
return c.generation
}
// supportsRPCV1 reports whether the live connection explicitly negotiated
// rpc-v1. Call repeats this check while capturing the sender under the same
// mutex, so this method is only a fast-path hint and cannot authorize a send by
// itself.
func (c *wsRPCClient) supportsRPCV1() bool {
if c == nil {
return false
}
c.mu.Lock()
defer c.mu.Unlock()
return c.sendFrame != nil && c.rpcV1Supported
}
// Call issues an RPC on any attached connection. Transport-level tests and
// callers that have their own negotiation contract use this directly.
func (c *wsRPCClient) Call(ctx context.Context, method string, serverTimeout time.Duration, reqBody, respBody any) (int, error) {
return c.call(ctx, method, serverTimeout, reqBody, respBody, false)
}
// CallIfRPCV1Supported issues an RPC only when the currently attached
// connection explicitly negotiated rpc-v1. The capability check and sender
// capture happen under the same mutex, so a reconnect cannot redirect a call
// authorized by the previous connection onto its replacement.
func (c *wsRPCClient) CallIfRPCV1Supported(ctx context.Context, method string, serverTimeout time.Duration, reqBody, respBody any) (int, error) {
return c.call(ctx, method, serverTimeout, reqBody, respBody, true)
}
// call blocks until the response, the per-request timeout, or ctx
// cancellation. reqBody is marshaled into the request envelope; on a 2xx
// response respBody (if non-nil) is unmarshaled from the response body. It
// returns the response status (0 when the call never reached the server) so the
// caller can distinguish transport failure (→ HTTP fallback) from a
// server-side error.
func (c *wsRPCClient) call(ctx context.Context, method string, serverTimeout time.Duration, reqBody, respBody any, requireRPCV1 bool) (int, error) {
if c == nil {
return 0, errWSRPCUnavailable
}
var rawReq json.RawMessage
if reqBody != nil {
b, err := json.Marshal(reqBody)
if err != nil {
return 0, fmt.Errorf("ws rpc: marshal request: %w", err)
}
rawReq = b
}
id := uuid.NewString()
frame, err := json.Marshal(protocol.Message{
Type: protocol.EventDaemonRPCRequest,
Payload: marshalRaw(protocol.RPCRequestPayload{
RequestID: id,
Method: method,
Body: rawReq,
TimeoutMs: serverTimeout.Milliseconds(),
}),
})
if err != nil {
return 0, fmt.Errorf("ws rpc: marshal frame: %w", err)
}
ch := make(chan protocol.RPCResponsePayload, 1)
c.mu.Lock()
if c.sendFrame == nil || (requireRPCV1 && !c.rpcV1Supported) {
c.mu.Unlock()
return 0, errWSRPCUnavailable
}
send := c.sendFrame
c.pending[id] = ch
c.mu.Unlock()
defer func() {
c.mu.Lock()
delete(c.pending, id)
c.mu.Unlock()
}()
item, err := send(frame)
if err != nil {
return 0, fmt.Errorf("ws rpc: send: %w", err)
}
// giveUp resolves an abandoned request. If the frame is still queued we
// cancel it so the writer never delivers it — a definitively-not-sent
// outcome that is safe to HTTP-fall-back. If the writer already began
// sending it, it may reach the server, so the outcome is uncertain and the
// caller must NOT fall back (that would double-claim, MUL-4257).
giveUp := func() error {
if item.cancel() {
return errWSRPCUnavailable
}
return errWSRPCUncertain
}
// Wait the server-side budget PLUS a grace margin: a claim that committed
// just before the server deadline must still report back before the daemon
// gives up and falls back to HTTP, or we would double-claim (MUL-4257).
timeout := serverTimeout + c.grace
if timeout <= 0 {
timeout = 5 * time.Second
}
timer := time.NewTimer(timeout)
defer timer.Stop()
select {
case resp, ok := <-ch:
if !ok {
// The connection detached. Whether the server saw this request
// depends on whether the frame had already left the writer, so let
// giveUp() decide (not-sent → safe fallback; sent → uncertain).
return 0, giveUp()
}
if resp.Status >= 200 && resp.Status < 300 {
if respBody != nil && len(resp.Body) > 0 {
if err := json.Unmarshal(resp.Body, respBody); err != nil {
return resp.Status, fmt.Errorf("ws rpc: decode response: %w", err)
}
}
return resp.Status, nil
}
msg := resp.Error
if msg == "" {
msg = fmt.Sprintf("ws rpc status %d", resp.Status)
}
return resp.Status, errors.New(msg)
case <-timer.C:
// The budget elapsed. If the frame is still queued behind a
// backpressured writer, cancel it so it is never delivered after we
// fall back (giveUp → not-sent). If it already left the writer, the
// outcome is uncertain and we must not fall back.
if err := giveUp(); errors.Is(err, errWSRPCUncertain) {
return 0, err
}
return 0, fmt.Errorf("ws rpc: timeout after %s: %w", timeout, errWSRPCUnavailable)
case <-ctx.Done():
item.cancel()
return 0, ctx.Err()
}
}
// deliver routes an inbound rpc_response frame to the waiting Call. The send
// happens under the mutex so it is serialized with attach(nil)'s close+delete:
// a channel present in pending is guaranteed not yet closed, so this never
// sends on a closed channel. Unknown request ids (already timed out / detached)
// are dropped.
func (c *wsRPCClient) deliver(resp protocol.RPCResponsePayload) {
if c == nil {
return
}
c.mu.Lock()
defer c.mu.Unlock()
ch, ok := c.pending[resp.RequestID]
if !ok {
return
}
select {
case ch <- resp:
default:
}
}
// ClaimTasksWSFirst is the WS-first claim policy (MUL-4257): it issues the
// tasks.claim RPC over the WS control connection when one is attached, and
// falls back to the HTTP claim endpoint on transport failures that are known not
// to have reached the server (no connection, write-buffer full, unsent timeout)
// or server error. A sent-frame disconnect/timeout is uncertain, so it is
// retried over HTTP only after a short safety window. The request/response bodies
// are identical to the HTTP endpoint so both transports are interchangeable.
// Wired into the claim poller as part of the poller cutover.
func (d *Daemon) ClaimTasksWSFirst(ctx context.Context, daemonID string, runtimeIDs []string, maxTasks int) ([]*Task, error) {
// Un-upgraded server without the batch route: a prior poll already learned
// this (via a 404), so go straight to the legacy per-runtime claim and skip
// the WS + batch attempts each cycle.
if d.batchClaimUnsupported.Load() {
return d.client.claimTasksLegacy(ctx, runtimeIDs, maxTasks)
}
bypassWSOnce := false
if retryAfterNanos := d.wsClaimHTTPFallbackAfter.Load(); retryAfterNanos > 0 {
retryAfter := time.Unix(0, retryAfterNanos)
now := time.Now()
if now.Before(retryAfter) {
d.logger.Debug("ws claim outcome uncertain; delaying http fallback until safety window elapses",
"retry_after", retryAfter.Sub(now).Round(time.Millisecond))
return nil, nil
}
if d.wsClaimHTTPFallbackAfter.CompareAndSwap(retryAfterNanos, 0) {
bypassWSOnce = true
d.logger.Debug("previous ws claim outcome uncertain; using http fallback for this claim cycle")
}
}
if !bypassWSOnce && d.wsRPC.supportsRPCV1() {
var resp struct {
Tasks []*Task `json:"tasks"`
}
// batchClaimRequestTimeout is the server-side execution budget; the
// daemon waits that plus the client's grace margin for the response.
_, err := d.wsRPC.CallIfRPCV1Supported(ctx, "tasks.claim", batchClaimRequestTimeout, map[string]any{
"daemon_id": daemonID,
"runtime_ids": runtimeIDs,
"max_tasks": maxTasks,
}, &resp)
if err == nil {
return resp.Tasks, nil
}
if errors.Is(err, errWSRPCUncertain) {
// The WS claim may have committed server-side; claiming the same
// free slots again over HTTP immediately would double-claim. Skip
// this cycle, then force one HTTP batch claim after the server-side
// execution budget plus response grace has elapsed. If the WS claim
// committed, the task is already dispatched and stale reclaim owns
// recovery; if it did not, HTTP regains liveness for the queued task.
delay := wsClaimUncertainFallbackDelay
if delay < 0 {
delay = 0
}
d.wsClaimHTTPFallbackAfter.Store(time.Now().Add(delay).UnixNano())
d.logger.Debug("ws claim outcome uncertain after disconnect; delaying http fallback", "retry_after", delay)
return nil, nil
}
d.logger.Debug("ws claim failed; falling back to http", "error", err)
}
tasks, err := d.client.ClaimTasks(ctx, daemonID, runtimeIDs, maxTasks)
if err == nil {
return tasks, nil
}
// Server has no batch route (404): freeze the old API contract by falling
// back to the legacy per-runtime claim loop, and remember it so we don't
// re-probe every cycle.
if isBatchClaimUnsupported(err) {
d.batchClaimUnsupported.Store(true)
d.logger.Info("batch claim route unsupported by server; using legacy per-runtime claim")
return d.client.claimTasksLegacy(ctx, runtimeIDs, maxTasks)
}
return nil, err
}