8e550b9785
This commit captures both the prior accumulated work-in-progress
(framework migration web/→svelte/, postgres storage, conversation
viewer, dashboard auth, OpenAPI spec, integration tests) AND today's
operational improvements layered on top. History wasn't checkpointed
incrementally; happy to split it via interactive rebase if a reviewer
wants smaller commits.
Today's changes (in addition to the older WIP):
1. Configurable upstream response-header timeout
- ANTHROPIC_RESPONSE_HEADER_TIMEOUT env (default 300s)
- Replaces hardcoded 300s in provider/anthropic.go that was firing
on opus + 1M-context + extended thinking non-streaming requests
- Files: internal/config/config.go, internal/provider/anthropic.go
2. Structured forward-error diagnostic logging
- When a forward to Anthropic fails, log a single key=value line
with request_id, model, stream, body_bytes, has_thinking,
anthropic_beta, query, elapsed, ctx_err — alongside the existing
human-readable error line for back-compat
- Files: internal/handler/handlers.go (logForwardFailure)
3. Full SSE protocol passthrough + Flusher fix
- handler/handlers.go: forward all SSE lines verbatim (event:, id:,
retry:, : comments, blank-line terminators), not only data:.
Previous code produced malformed SSE for strict parsers.
- middleware/logging.go: explicit Flush() method on responseWriter.
Embedding http.ResponseWriter (interface) does not auto-promote
Flush(), so every w.(http.Flusher) check in the streaming
handler was returning ok=false and SSE writes buffered in net/http
until the body closed.
4. Non-streaming → streaming demotion (feature-flagged)
- ANTHROPIC_DEMOTE_NONSTREAMING env (default false)
- When enabled and the routed provider is anthropic, force stream=true
upstream for clients that asked for stream=false. Receive SSE,
accumulate via accumulateSSEToMessage (handles text, tool_use with
partial_json reassembly, thinking, signature, citations_delta,
usage merge), and synthesize a single non-streaming JSON response.
- Eliminates the ResponseHeaderTimeout class of failure entirely.
- Body rewrite uses json.Decoder + UseNumber() to preserve integer
precision in unknown nested fields (tool inputs from prior turns).
- Files: internal/config/config.go, internal/handler/handlers.go,
cmd/proxy/main.go, cmd/proxy/main_test.go
5. Live operational state: /livez gauge + graceful drain
- New internal/runtime package: atomic in-flight counter + draining flag
- New middleware/inflight.go: increments runtime gauge, applied to
/v1/* subrouter so Messages, ChatCompletions, and ProxyPassthrough
are all counted
- /v1/* moved to a gorilla/mux subrouter so the InFlight middleware
applies surgically; /health, /livez, /openapi.* remain on parent
router (unauthenticated, uncounted)
- Health handler returns 503 draining when runtime.IsDraining() is
true, so Traefik stops routing to a slot before drain begins
- New /livez handler returns {status, in_flight, draining, timestamp}
- SIGTERM handler in main.go: SetDraining(true), poll for in_flight==0
with 32-min ceiling and 1s tick (logs every 10s), then srv.Shutdown
- Auth bypass list extended with /livez
- Files: internal/runtime/runtime.go (new),
internal/middleware/inflight.go (new),
internal/middleware/auth.go,
internal/handler/handlers.go (Health, Livez, runtime import),
cmd/proxy/main.go (subrouter, drain loop)
6. OpenAPI spec updates
- Document Health 503 response and new DrainingResponse schema
- Add /livez path with LivezResponse schema
- Files: internal/handler/openapi.go
Verified: go build ./... clean, go test ./... all pass, go vet clean.
Three rounds of codex peer review across changes 1-5; all feedback
addressed (citations_delta, json.Number precision, drain-loop logging
via lastLog timestamp, PathPrefix tightened to "/v1/").
34 lines
1.4 KiB
Go
34 lines
1.4 KiB
Go
// Package runtime exposes process-level operational state shared between
|
|
// HTTP middleware, handlers, and the shutdown loop in main: a live in-flight
|
|
// request gauge for /livez, and a draining flag that flips on SIGTERM so
|
|
// /health goes non-ready before we wait for in-flight requests to drain.
|
|
//
|
|
// All state is package-level + atomic so callers don't need to plumb a struct
|
|
// through every middleware/handler. The gauge is decremented in a deferred
|
|
// call regardless of panics, so a misbehaving handler can't strand the count.
|
|
package runtime
|
|
|
|
import "sync/atomic"
|
|
|
|
var (
|
|
inFlight atomic.Int64
|
|
draining atomic.Bool
|
|
)
|
|
|
|
// IncInFlight is called when a tracked request begins. Returns the new count.
|
|
func IncInFlight() int64 { return inFlight.Add(1) }
|
|
|
|
// DecInFlight is called when a tracked request completes. Returns the new count.
|
|
func DecInFlight() int64 { return inFlight.Add(-1) }
|
|
|
|
// InFlight returns the current number of tracked requests in progress.
|
|
func InFlight() int64 { return inFlight.Load() }
|
|
|
|
// IsDraining reports whether the process is shutting down. Used by /health
|
|
// to advertise non-ready state so Traefik (or any LB doing health-based
|
|
// routing) stops sending new requests to this slot before drain begins.
|
|
func IsDraining() bool { return draining.Load() }
|
|
|
|
// SetDraining flips the draining flag. Idempotent — safe to call from a
|
|
// signal handler.
|
|
func SetDraining(v bool) { draining.Store(v) }
|