Merge pull request #12 from frappe/multi_queue

feat: Adaptive queueing of requests for gthread

docs/design/gthread-slow-request-isolation.md

@@ -0,0 +1,272 @@
+# Design: Slow-request isolation for the gthread worker (predictive dual-queue)
+
+Status: proposal / draft
+Author: (ankush)
+Scope: `gunicorn/workers/gthread.py`, `gunicorn/config.py`
+
+## 1. Problem
+
+The `gthread` worker runs synchronous WSGI applications on a single
+`ThreadPoolExecutor` sized to `cfg.threads` (`gthread.py:95-97`). Every accepted
+connection is submitted to that one pool (`enqueue_req`, `gthread.py:117-121`).
+Because the pool has a fixed number of threads and an unbounded work queue, a
+flood of slow requests occupies every thread and all fast requests starve behind
+them in the queue — head-of-line blocking.
+
+Goal: **route requests that are predicted to be slow into a separate, dedicated
+lane so they can never occupy the threads reserved for fast requests, even under
+a flood.** Fast requests go to a fast lane; slow requests go to a slow lane. The
+slow lane may help drain fast work when its own queue is empty, but the fast
+lane never touches slow work.
+
+This supersedes the earlier "demotion-only" proposal, which could not stop slow
+work from entering the fast pool and therefore could not survive a flood.
+
+## 2. Why prediction is required (and its hard limit)
+
+You cannot preempt a running Python thread executing WSGI code (`gthread.py:352`):
+once a slow request is on a thread, that thread is committed until the app
+returns. So isolation has to happen **before** a request is handed to a worker —
+i.e. at routing time. That means we must decide "fast or slow" from the request
+*before* running it.
+
+The only information available pre-execution is the request itself (method,
+path, headers) plus what we have learned from prior requests. So the design is:
+
+1. A **predictor** that, given a request's route, answers "slow?" using learned
+   per-route timing statistics (plus optional operator-seeded patterns).
+2. **Routing at accept time** based on that prediction, into one of two pools.
+3. **Learning**: every completed request — and any request that crosses the
+   slow threshold mid-flight — updates the predictor, so a slow route is
+   recognized after its first occurrence(s) and all subsequent traffic to it is
+   routed to the slow lane.
+
+Hard limit to state up front: a route that has **never been seen** cannot be
+predicted slow on its very first request(s); those first occurrences run in the
+fast lane until learning kicks in. We minimize this window (§5.4) and let
+operators pre-seed known-slow routes (§5.1). For repeated/flooding slow routes —
+the actual failure mode — prediction is effective after the first sample.
+
+## 3. Architecture overview
+
+```
+                          ┌─────────────────────────────────────────┐
+   listener ──accept──▶   │  main loop: poller-driven classification  │
+                          │  peek request line ▶ predictor.is_slow?   │
+                          └───────────────┬───────────────┬───────────┘
+                                          │ fast          │ slow
+                                          ▼               ▼
+                                   ┌───────────┐   ┌───────────┐
+                                   │ fast_pool │   │ slow_pool │
+                                   │ F threads │   │ S threads │
+                                   └─────┬─────┘   └─────┬─────┘
+                                         └───────┬───────┘
+                              on completion: predictor.update(route, duration)
+```
+
+- **Fast lane**: a `ThreadPoolExecutor` of `F = ceil(cfg.threads / 2)` threads.
+  Only ever runs fast-classified work.
+- **Slow lane**: a separate `ThreadPoolExecutor` of `S = cfg.threads // 2`
+  threads. Only ever runs slow-classified work.
+- Total OS threads per worker stays at `cfg.threads` — adaptive-queueing mode splits
+  the existing budget, it does not expand it. `cfg.threads` must be at least 2
+  for the split to be meaningful; otherwise the worker logs a warning and runs
+  with a single pool.
+- Both lanes share the existing `worker_connections` admission like today; the
+  slow lane is not separately bounded.
+
+### Why two plain pools (and not a custom dual-queue scheduler)
+
+An earlier revision used a single custom scheduler with two queues and
+one-directional **work stealing** (idle slow threads draining the fast queue).
+Two independent `ThreadPoolExecutor`s are dramatically simpler and rely on
+well-tested stdlib machinery. The one capability given up is work stealing: the
+`S` slow threads sit idle when there is no slow work, even if fast work is
+queued. For the common case (`S` small, e.g. 1) this is a negligible amount of
+parked capacity, and the simplicity is worth it. If maximizing throughput under
+pure-fast load ever matters more than simplicity, the custom scheduler can be
+reintroduced behind the same `enqueue_req` interface without touching routing.
+
+## 4. Routing point: classify before threading
+
+Today, parsing happens inside the worker thread (`handle` → `next(conn.parser)`,
+`gthread.py:295`), which is too late — the request is already on a thread. We
+move *classification only* (not full parsing) into the main loop.
+
+### 4.1 Restructured connection lifecycle
+
+Both freshly accepted connections and keepalive connections flow through one
+poller-driven classification step (this unifies `accept`/`reuse_connection` and
+also moves slow-client header reads off the worker threads — a side benefit
+against slowloris):
+
+1. `accept` (`gthread.py:123`): accept socket, create `TConn`, set non-blocking,
+   register it in the poller for `EVENT_READ` with a `classify_and_dispatch`
+   callback. **Do not submit to any pool yet.** `nr_conns += 1`.
+2. When the socket becomes readable, `classify_and_dispatch(conn)`:
+   - **Peek** the buffered bytes with `recv(n, socket.MSG_PEEK)` (plaintext) —
+     this reads without consuming, so the worker's parser still sees the full
+     byte stream unchanged. No parser changes required.
+   - Parse just the request line (`METHOD SP PATH SP VERSION CRLF`) from the
+     peeked buffer. If the line has not fully arrived yet, return and wait for
+     the next readable event (bounded by the existing keepalive/header timeout so
+     a stalled client is eventually closed, not left forever).
+
+   > **Why peek the request line, not fully read/parse the request here?**
+   > Classification only needs method + path. Doing a *full* read/parse of the
+   > request in the main loop is actively harmful: the main loop is a single
+   > thread serving every connection (accepts, keepalive, the poller). A blocking
+   > full read lets one slow client — slowloris, slow network, or a large/chunked
+   > body — stall the **entire worker**, which is strictly worse than the
+   > thread-pool starvation we are fixing (there is no pool to absorb it).
+   > Peeking only inspects already-buffered bytes and defers to the poller if the
+   > line is incomplete, so it never blocks. It also avoids having to read the
+   > body in the main loop (WSGI streams `wsgi.input` lazily) and keeps header
+   > parsing, parse-error responses (400/414), and `wsgi.input` wiring in the
+   > worker where they already live.
+   - Compute `route_key` (default: `METHOD + " " + path`, query string stripped;
+     overridable via hook, §5.1).
+   - `slow = predictor.is_slow(route_key)` (or matches a seeded slow pattern).
+   - Unregister the socket from the poller and submit the connection to the
+     **slow** pool if `slow` else the **fast** pool.
+3. The worker's `handle`/`handle_request` run unchanged. On completion, the
+   measured `request_time` (already computed at `gthread.py:362`) is fed to
+   `predictor.update(route_key, duration)`.
+4. Keepalive: after a kept-alive request, re-register the connection in the
+   poller with the same `classify_and_dispatch` callback (instead of the old
+   `reuse_connection`), so the *next* request on the connection is re-classified
+   independently (it may hit a different route).
+
+### 4.2 SSL connections
+
+Plaintext peek does not work through TLS — the request line is encrypted until
+the handshake completes. For SSL connections in this first cut:
+
+- They cannot be pre-classified at the socket level, so they default to the
+  **fast** lane and rely on mid-flight + completion learning (§5.4) — meaning an
+  SSL-only deployment does not get full flood protection.
+- Note in docs that the common production layout terminates TLS upstream (e.g.
+  nginx) so gunicorn sees plaintext and gets full protection.
+- **Phase 2** (deferred): drive a non-blocking TLS handshake from the poller and
+  buffer the decrypted request line (feeding it back via `Unreader.unread`,
+  `unreader.py:51`) to classify SSL the same way.
+
+## 5. Components
+
+### 5.1 Config (`gunicorn/config.py`)
+
+New settings, mirroring `WorkerThreads` (`config.py:697`):
+
+- `enable_adaptive_queueing` — boolean; when true, the `gthread` worker splits its
+  `cfg.threads` budget between a fast and a slow lane and routes by prediction.
+  Default `False` (single pool, today's behavior). Requires `cfg.threads >= 2`;
+  otherwise the worker logs a warning and falls back to the single pool.
+- `slow_request_threshold` — float seconds; a route whose learned timing meets/
+  exceeds this is "slow". Default `1.0`. Only consulted when `enable_adaptive_queueing` is
+  enabled.
+
+A `slow_route_key` hook to customize the route key (e.g. collapse
+`/users/<id>`) is a possible future addition; the default key is method + path
+with the query string stripped.
+
+### 5.2 Two thread pools
+
+`init_process` builds two plain `ThreadPoolExecutor`s when `enable_adaptive_queueing` is
+enabled — `fast_pool` with `F = ceil(cfg.threads / 2)` workers and `slow_pool`
+with `S = cfg.threads // 2` workers — and falls back to the single
+`get_thread_pool()` executor when it is disabled. `enqueue_req(conn, slow)`
+submits to the matching pool; both produce ordinary
+`concurrent.futures.Future`s, so `_wrap_future`, `add_done_callback`,
+`self.futures` tracking, and `futures.wait` all keep working unchanged.
+
+- The slow lane is not separately bounded; its executor's internal queue is
+  capped indirectly by the worker's existing `worker_connections` admission,
+  the same as the single-pool path today.
+- Shutdown drains both pools via a `_shutdown_pools` helper, replacing the
+  single `tpool.shutdown` calls; the `graceful_timeout` `futures.wait` is
+  unchanged.
+
+### 5.3 Predictor
+
+A small, self-contained, thread-safe object:
+
+- State: bounded LRU map `route_key -> {ewma_seconds, samples, last_seen}`.
+  Bounding caps memory under high route cardinality.
+- `update(route_key, duration)`: EWMA with decay so a route that becomes fast
+  again eventually returns to the fast lane (avoids permanent misclassification
+  after a one-off slow spike). Called on every completion.
+- `is_slow(route_key)`: `True` if its `ewma_seconds >= slow_request_threshold`.
+  Unknown routes ⇒ `False` (fast) by default.
+- Optional hysteresis (separate promote/demote thresholds) to avoid flapping
+  around the boundary.
+
+### 5.4 Learning signals
+
+1. **Completion (primary)**: feed `request_time` (`gthread.py:362`) into
+   `predictor.update`. After a slow route's first request completes, it is known.
+2. **Mid-flight observation (catches simultaneous first-bursts)**: the main loop
+   already sweeps `self.futures` for the hard timeout (`gthread.py:245-250`). In
+   that sweep, for any in-flight request whose elapsed time exceeds
+   `slow_request_threshold`, call `predictor.update` with that elapsed time
+   *immediately* (do not wait for completion, and do not move the running
+   request — we can't). This shortens the learning window when many requests to a
+   brand-new slow route arrive at once: subsequent ones in the burst route to the
+   slow lane after one threshold interval instead of after a full slow request.
+
+## 6. Behavior under load (the cases that matter)
+
+- **Flood of a previously-seen slow route**: every such request is routed to
+  the slow pool. The `F` fast threads are never given this work and keep
+  serving fast traffic at full capacity. Excess slow requests sit in the slow
+  pool's queue, gated overall by `worker_connections`.
+- **Flood of a never-seen slow route**: the first occurrence(s) run in the fast
+  lane; mid-flight learning (§5.4.2) flips the route to slow after one threshold
+  interval, so the flood is contained quickly.
+- **Mixed fast traffic, idle slow lane**: the `S` slow threads stay parked (no
+  work stealing in this design — see §3), so fast throughput is `F`, not
+  `F + S`. This is the cost of splitting a fixed `cfg.threads` budget.
+- **Misprediction (route marked slow but now fast)**: handled gracefully — it
+  runs in the slow lane, and EWMA decay restores it to the fast lane over time.
+
+## 7. Implementation checklist (touch points)
+
+Implemented:
+
+- `config.py` — `enable_adaptive_queueing`, `slow_request_threshold`, plus
+  `validate_pos_float`.
+- `gthread.py` `init_process`/`get_thread_pool` — build `fast_pool` and
+  `slow_pool` (split from `cfg.threads`) when `enable_adaptive_queueing` is on, or the
+  single legacy pool when off; `_shutdown_pools`.
+- `gthread.py` `enqueue_req` — route to the matching pool.
+- `gthread.py` `accept`/`park_for_request`/`classify_and_dispatch`/
+  `_peek_request_line`/`_route_key` — poller-driven request-line peek + routing.
+- `gthread.py` `finish_request` — `predictor.update`, routing-aware keepalive
+  re-park.
+- `gthread.py` run-loop sweep — mid-flight learning.
+- `gthread_routing.py` — `SlowRoutePredictor`.
+
+## 8. Backward compatibility
+
+- `enable_adaptive_queueing = False` (the default) ⇒ feature off: single pool, no
+  classification — byte-for-byte current behavior.
+- Hard per-request timeout (`gthread.py:243-250`) preserved unchanged; this adds
+  a softer, non-fatal classification on top.
+- Worker `handle`/`handle_request`, keepalive semantics, and the
+  future/`finish_request` contract are preserved (MSG_PEEK leaves the byte
+  stream intact, so the parser is untouched).
+
+## 9. Test plan
+
+- **Predictor unit**: unknown ⇒ fast; after `update` with a slow duration ⇒
+  slow; EWMA decay restores fast; seeded patterns are slow from first call; LRU
+  bound holds under many keys.
+- **Routing unit**: `classify_and_dispatch` extracts the right `route_key` from
+  partial vs complete peeked buffers; incomplete line defers; complete line
+  dispatches to the expected lane.
+- **Integration — flood isolation**: app with a known-slow route flooded
+  concurrently; assert fast-route latency stays low and slow requests never
+  occupy fast workers.
+- **Integration — cold start**: never-seen slow route burst ⇒ confirm the lane
+  flips to slow within ~one threshold interval via mid-flight learning.
+- **Regression**: `enable_adaptive_queueing = False` ⇒ current behavior; keepalive, SSL, and
+  graceful shutdown paths still pass existing tests.

gunicorn/config.py

@@ -365,6 +365,13 @@ def validate_pos_int(val):
     return val
 
 
+def validate_pos_float(val):
+    val = float(val)
+    if val < 0:
+        raise ValueError("Value must be positive: %s" % val)
+    return val
+
+
 def validate_ssl_version(val):
     if val != SSLVersion.default:
         sys.stderr.write("Warning: option `ssl_version` is deprecated and it is ignored. Use ssl_context instead.\n")
@@ -737,6 +744,61 @@ class WorkerConnections(Setting):
         """
 
 
+class EnableAdaptiveQueueing(Setting):
+    name = "enable_adaptive_queueing"
+    section = "Worker Processes"
+    cli = ["--enable-adaptive-queueing"]
+    validator = validate_bool
+    action = "store_true"
+    type = bool
+    default = validate_bool(
+        os.environ.get("GUNICORN_ENABLE_ADAPTIVE_QUEUEING", "false"))
+    desc = """\
+        Enable adaptive multi-queue routing in the ``gthread`` worker.
+
+        Can also be enabled by setting the ``GUNICORN_ENABLE_ADAPTIVE_QUEUEING``
+        environment variable to ``true``.
+
+        When enabled, the worker splits its :ref:`threads` roughly evenly into
+        two lanes — a *fast* lane and a *slow* lane — and routes each request
+        to one of them by predicting, from previously observed timings of the
+        same route (method + path), whether it will exceed
+        :ref:`slow-request-threshold`. Slow-predicted requests go to the slow
+        lane so they can never starve the fast lane, even under a flood of
+        slow requests.
+
+        Requires :ref:`threads` to be at least 2.
+
+        This setting only affects the ``gthread`` worker type.
+
+        .. versionadded:: 23.1.0
+        """
+
+
+class SlowRequestThreshold(Setting):
+    name = "slow_request_threshold"
+    section = "Worker Processes"
+    cli = ["--slow-request-threshold"]
+    meta = "FLOAT"
+    validator = validate_pos_float
+    type = float
+    default = 5.0
+    desc = """\
+        Processing time (in seconds) above which a request route is treated as
+        "slow" by the ``gthread`` worker when :ref:`enable-adaptive-queueing`
+        is enabled.
+
+        A route is learned as slow once it has been observed exceeding this
+        threshold (either on completion or while still running); its timing
+        decays back below the threshold if it becomes fast again.
+
+        Only used by the ``gthread`` worker when :ref:`enable-adaptive-queueing`
+        is enabled.
+
+        .. versionadded:: 23.1.0
+        """
+
+
 class MaxRequests(Setting):
     name = "max_requests"
     section = "Worker Processes"

gunicorn/workers/gthread.py

@@ -25,11 +25,15 @@ from functools import partial
 from threading import RLock
 
 from . import base
+from .gthread_routing import SlowRoutePredictor
 from .. import http
 from .. import util
 from .. import sock
 from ..http import wsgi
 
+# how many bytes to peek when classifying a request by its request line
+REQUEST_LINE_PEEK = 8192
+
 
 class TConn:
 
@@ -41,6 +45,9 @@ class TConn:
 
         self.timeout = None
         self.parser = None
+        # route key (method + path), set by the worker when request routing is
+        # enabled; used to predict and learn slow routes
+        self.route_key = None
 
         # set the socket to non blocking
         self.sock.setblocking(False)
@@ -69,14 +76,23 @@ class ThreadWorker(base.Worker):
         super().__init__(*args, **kwargs)
         self.worker_connections = self.cfg.worker_connections
         self.max_keepalived = self.cfg.worker_connections - self.cfg.threads
-        # initialise the pool
+        # request routing: when adaptive queueing is enabled, the configured
+        # threads are split into a fast lane and a slow lane so slow requests
+        # cannot starve fast ones
+        self.routing_enabled = (
+            self.cfg.enable_adaptive_queueing and self.cfg.threads >= 2)
+        self.slow_threshold = self.cfg.slow_request_threshold
+        # initialise the pool(s): a single pool when routing is disabled, or a
+        # separate fast (``self.tpool``) and slow pool when it is enabled
         self.tpool = None
+        self.slow_pool = None
         self.poller = None
         self.shutdown_event = os.eventfd(0)
         self._lock = None
         self.futures = deque()
         self._keep = deque()
         self.nr_conns = 0
+        self.predictor = None
 
     @classmethod
     def check_config(cls, cfg, log):
@@ -86,8 +102,21 @@ class ThreadWorker(base.Worker):
             log.warning("No keepalived connections can be handled. " +
                         "Check the number of worker connections and threads.")
 
+        if cfg.enable_adaptive_queueing and cfg.threads < 2:
+            log.warning("enable_adaptive_queueing requires at least 2 threads; "
+                        "running with a single pool.")
+
     def init_process(self):
-        self.tpool = self.get_thread_pool()
+        if self.routing_enabled:
+            # split the configured threads roughly evenly between the two
+            # lanes; the fast lane gets the extra thread when threads is odd
+            slow = self.cfg.threads // 2
+            fast = self.cfg.threads - slow
+            self.tpool = futures.ThreadPoolExecutor(max_workers=fast)
+            self.slow_pool = futures.ThreadPoolExecutor(max_workers=slow)
+            self.predictor = SlowRoutePredictor(self.slow_threshold)
+        else:
+            self.tpool = self.get_thread_pool()
         self.poller = selectors.DefaultSelector()
         self._lock = RLock()
         super().init_process()
@@ -96,6 +125,11 @@ class ThreadWorker(base.Worker):
         """Override this method to customize how the thread pool is created"""
         return futures.ThreadPoolExecutor(max_workers=self.cfg.threads)
 
+    def _shutdown_pools(self, wait):
+        for pool in (self.tpool, self.slow_pool):
+            if pool is not None:
+                pool.shutdown(wait)
+
     def handle_exit(self, sig, frame):
         self.alive = False
         os.eventfd_write(self.shutdown_event, 1)
@@ -104,21 +138,26 @@ class ThreadWorker(base.Worker):
         self.alive = False
         # worker_int callback
         self.cfg.worker_int(self)
-        self.tpool.shutdown(False)
+        self._shutdown_pools(False)
         time.sleep(0.1)
         sys.exit(0)
 
-    def _wrap_future(self, fs, conn):
+    def _wrap_future(self, fs, conn, slow=False):
         fs.conn = conn
-        fs._request_timeout = time.monotonic() + self.cfg.timeout
+        fs.slow = slow
+        fs._start_time = time.monotonic()
+        fs._request_timeout = fs._start_time + self.cfg.timeout
+        fs._observed_slow = False
         self.futures.append(fs)
         fs.add_done_callback(self.finish_request)
 
-    def enqueue_req(self, conn):
+    def enqueue_req(self, conn, slow=False):
         conn.init()
-        # submit the connection to a worker
+        if self.routing_enabled and slow:
-        fs = self.tpool.submit(self.handle, conn)
+            fs = self.slow_pool.submit(self.handle, conn)
-        self._wrap_future(fs, conn)
+        else:
+            fs = self.tpool.submit(self.handle, conn)
+        self._wrap_future(fs, conn, slow=slow)
 
     def accept(self, server, listener):
         try:
@@ -127,13 +166,99 @@ class ThreadWorker(base.Worker):
             conn = TConn(self.cfg, sock, client, server)
 
             self.nr_conns += 1
-            # enqueue the job
+            if self.routing_enabled and not self.cfg.is_ssl:
-            self.enqueue_req(conn)
+                # park the connection until its request line is readable, then
+                # classify and route it to the fast or slow lane
+                self.park_for_request(conn)
+            else:
+                # legacy single-lane path (also used for SSL, whose request
+                # line cannot be peeked before the TLS handshake)
+                self.enqueue_req(conn)
         except OSError as e:
             if e.errno not in (errno.EAGAIN, errno.ECONNABORTED,
                                errno.EWOULDBLOCK):
                 raise
 
+    def park_for_request(self, conn):
+        """Register a connection in the poller until its request line arrives."""
+        conn.sock.setblocking(False)
+        conn.set_timeout()
+        with self._lock:
+            self._keep.append(conn)
+            self.poller.register(conn.sock, selectors.EVENT_READ,
+                                 partial(self.classify_and_dispatch, conn))
+
+    def classify_and_dispatch(self, conn, client=None):
+        """Peek the request line, predict the lane, and enqueue the request."""
+        line, closed, complete = self._peek_request_line(conn)
+        if not closed and not complete:
+            # request line has not fully arrived yet; keep waiting. Stalled
+            # clients are reaped by murder_keepalived via the connection timeout.
+            return
+
+        with self._lock:
+            try:
+                # remove the connection from the parked set
+                self._keep.remove(conn)
+            except ValueError:
+                # already handled (e.g. by murder_keepalived); nothing to do
+                return
+            try:
+                self.poller.unregister(conn.sock)
+            except (KeyError, OSError, ValueError):
+                pass
+
+        if closed:
+            self.nr_conns -= 1
+            conn.close()
+            return
+
+        conn.route_key = self._route_key(line)
+        slow = self.predictor.is_slow(conn.route_key)
+        self.enqueue_req(conn, slow=slow)
+
+    def _peek_request_line(self, conn):
+        """Return ``(line, closed, complete)`` for the connection's request line.
+
+        ``line`` is the request line bytes (without CRLF) once available,
+        ``closed`` is True if the peer closed the connection, and ``complete``
+        is True once we should stop waiting for more data.
+        """
+        try:
+            data = conn.sock.recv(REQUEST_LINE_PEEK, socket.MSG_PEEK)
+        except (BlockingIOError, InterruptedError):
+            return None, False, False
+        except OSError:
+            return None, True, False
+
+        if data == b"":
+            # peer closed the connection before sending a request
+            return None, True, False
+
+        idx = data.find(b"\r\n")
+        if idx == -1:
+            if len(data) >= REQUEST_LINE_PEEK:
+                # request line longer than our peek window; stop classifying and
+                # let the worker's parser deal with (or reject) it
+                return None, False, True
+            return None, False, False
+        return data[:idx], False, True
+
+    @staticmethod
+    def _route_key(line):
+        """Build a route key (``"METHOD /path"``) from a raw request line."""
+        if not line:
+            return None
+        parts = line.split(b" ")
+        if len(parts) < 2:
+            return None
+        try:
+            method = parts[0].decode("latin1")
+            path = parts[1].split(b"?", 1)[0].decode("latin1")
+        except UnicodeDecodeError:
+            return None
+        return method + " " + path
+
     def reuse_connection(self, conn, client):
         with self._lock:
             # unregister the client from the poller
@@ -248,8 +373,16 @@ class ThreadWorker(base.Worker):
                     self.alive = False
                     self.log.error("A request timed out. Exiting.")
                     faulthandler.dump_traceback()
+                elif (self.routing_enabled and not fut._observed_slow
+                        and not fut.slow
+                        and current_time - fut._start_time > self.slow_threshold):
+                    # an in-flight fast-lane request crossed the threshold; learn
+                    # the route as slow now so the rest of a burst is rerouted
+                    # without waiting for this request to finish
+                    self.predictor.observe_slow(fut.conn.route_key)
+                    fut._observed_slow = True
 
-        self.tpool.shutdown(False)
+        self._shutdown_pools(False)
         self.poller.close()
 
         for s in self.sockets:
@@ -263,22 +396,32 @@ class ThreadWorker(base.Worker):
             fs.conn.close()
             return
 
+        # feed the observed processing time back to the predictor so the route
+        # is learned (or unlearned) as slow
+        if self.routing_enabled and fs.conn.route_key:
+            self.predictor.update(fs.conn.route_key,
+                                  time.monotonic() - fs._start_time)
+
         try:
             (keepalive, conn) = fs.result()
             # if the connection should be kept alived add it
             # to the eventloop and record it
             if keepalive and self.alive:
-                # flag the socket as non blocked
+                if self.routing_enabled and not self.cfg.is_ssl:
-                conn.sock.setblocking(False)
+                    # re-classify the next request on this connection
+                    self.park_for_request(conn)
+                else:
+                    # flag the socket as non blocked
+                    conn.sock.setblocking(False)
 
-                # register the connection
+                    # register the connection
-                conn.set_timeout()
+                    conn.set_timeout()
-                with self._lock:
+                    with self._lock:
-                    self._keep.append(conn)
+                        self._keep.append(conn)
 
-                    # add the socket to the event loop
+                        # add the socket to the event loop
-                    self.poller.register(conn.sock, selectors.EVENT_READ,
+                        self.poller.register(conn.sock, selectors.EVENT_READ,
-                                         partial(self.reuse_connection, conn))
+                                             partial(self.reuse_connection, conn))
             else:
                 self.nr_conns -= 1
                 conn.close()

gunicorn/workers/gthread_routing.py

@@ -0,0 +1,75 @@
+#
+# This file is part of gunicorn released under the MIT license.
+# See the NOTICE for more information.
+
+"""Slow-route prediction for the gthread worker.
+
+The :class:`SlowRoutePredictor` decides, before a request is handed to a
+worker, whether its route is expected to be slow, based on previously observed
+timings of the same route (method + path). The gthread worker uses this to
+route slow requests to a dedicated thread pool so they cannot starve fast
+requests.
+"""
+
+import threading
+from collections import OrderedDict
+
+
+class SlowRoutePredictor:
+    """Predicts whether a route (method + path) is slow.
+
+    Timings are tracked per route as an exponentially weighted moving average
+    (EWMA) so that a route which becomes fast again decays back below the
+    threshold. The table is bounded (LRU) to cap memory under high route
+    cardinality.
+    """
+
+    def __init__(self, threshold, max_entries=1024, alpha=0.3):
+        self.threshold = threshold
+        self.alpha = alpha
+        self.max_entries = max_entries
+        self._stats = OrderedDict()
+        self._lock = threading.Lock()
+
+    def is_slow(self, key):
+        if not key:
+            return False
+        with self._lock:
+            ewma = self._stats.get(key)
+            if ewma is None:
+                return False
+            self._stats.move_to_end(key)
+            return ewma >= self.threshold
+
+    def update(self, key, duration):
+        """Record an observed processing ``duration`` (seconds) for ``key``."""
+        if not key:
+            return
+        with self._lock:
+            ewma = self._stats.get(key)
+            if ewma is None:
+                ewma = duration
+            else:
+                ewma = (1 - self.alpha) * ewma + self.alpha * duration
+            self._stats[key] = ewma
+            self._stats.move_to_end(key)
+            self._evict()
+
+    def observe_slow(self, key):
+        """Mark ``key`` slow now, before its request has finished.
+
+        Used when an in-flight request crosses the threshold, so the rest of a
+        simultaneous burst to a never-seen slow route is routed to the slow lane
+        without waiting for the first request to complete.
+        """
+        if not key:
+            return
+        with self._lock:
+            cur = self._stats.get(key, 0.0)
+            self._stats[key] = max(cur, self.threshold)
+            self._stats.move_to_end(key)
+            self._evict()
+
+    def _evict(self):
+        while len(self._stats) > self.max_entries:
+            self._stats.popitem(last=False)

tests/workers/test_gthread.py

@@ -0,0 +1,97 @@
+#
+# This file is part of gunicorn released under the MIT license.
+# See the NOTICE for more information.
+
+import socket
+import types
+
+from gunicorn.workers.gthread import ThreadWorker, REQUEST_LINE_PEEK
+
+
+# _route_key is a staticmethod, so it can be exercised directly.
+
+def test_route_key_basic():
+    assert ThreadWorker._route_key(b"GET /index HTTP/1.1") == "GET /index"
+
+
+def test_route_key_strips_query_string():
+    assert ThreadWorker._route_key(
+        b"GET /search?q=hello&p=2 HTTP/1.1") == "GET /search"
+
+
+def test_route_key_post():
+    assert ThreadWorker._route_key(
+        b"POST /reports/generate HTTP/1.0") == "POST /reports/generate"
+
+
+def test_route_key_malformed():
+    assert ThreadWorker._route_key(b"") is None
+    assert ThreadWorker._route_key(b"GARBAGE") is None
+    assert ThreadWorker._route_key(None) is None
+
+
+def _peek(conn):
+    # _peek_request_line only touches conn.sock, so we can bind it to a stub
+    return ThreadWorker._peek_request_line(object(), conn)
+
+
+def test_peek_complete_request_line():
+    a, b = socket.socketpair()
+    try:
+        a.setblocking(False)
+        b.sendall(b"GET /x HTTP/1.1\r\nHost: y\r\n\r\n")
+        conn = types.SimpleNamespace(sock=a)
+        line, closed, complete = _peek(conn)
+        assert line == b"GET /x HTTP/1.1"
+        assert closed is False
+        assert complete is True
+        # MSG_PEEK must leave the bytes in the buffer for the parser
+        assert a.recv(5) == b"GET /"
+    finally:
+        a.close()
+        b.close()
+
+
+def test_peek_incomplete_request_line_waits():
+    a, b = socket.socketpair()
+    try:
+        a.setblocking(False)
+        b.sendall(b"GET /x HTT")  # no CRLF yet
+        conn = types.SimpleNamespace(sock=a)
+        line, closed, complete = _peek(conn)
+        assert line is None
+        assert closed is False
+        assert complete is False
+    finally:
+        a.close()
+        b.close()
+
+
+def test_peek_no_data_yet():
+    a, b = socket.socketpair()
+    try:
+        a.setblocking(False)
+        conn = types.SimpleNamespace(sock=a)
+        line, closed, complete = _peek(conn)
+        # nothing buffered: not closed, not complete -> keep waiting
+        assert (line, closed, complete) == (None, False, False)
+    finally:
+        a.close()
+        b.close()
+
+
+def test_peek_peer_closed():
+    a, b = socket.socketpair()
+    a.setblocking(False)
+    b.close()
+    try:
+        conn = types.SimpleNamespace(sock=a)
+        line, closed, complete = _peek(conn)
+        assert closed is True
+    finally:
+        a.close()
+
+
+def test_peek_window_constant_is_reasonable():
+    # a sanity bound so request lines fit comfortably in one peek
+    assert REQUEST_LINE_PEEK >= 8192

tests/workers/test_gthread_routing.py

@@ -0,0 +1,47 @@
+#
+# This file is part of gunicorn released under the MIT license.
+# See the NOTICE for more information.
+
+from gunicorn.workers.gthread_routing import SlowRoutePredictor
+
+
+def test_predictor_unknown_route_is_fast():
+    p = SlowRoutePredictor(threshold=1.0)
+    assert p.is_slow("GET /") is False
+
+
+def test_predictor_learns_slow_route_on_update():
+    p = SlowRoutePredictor(threshold=1.0, alpha=1.0)
+    p.update("GET /slow", 5.0)
+    assert p.is_slow("GET /slow") is True
+    assert p.is_slow("GET /fast") is False
+
+
+def test_predictor_ewma_decays_back_to_fast():
+    p = SlowRoutePredictor(threshold=1.0, alpha=0.5)
+    p.update("GET /x", 5.0)
+    assert p.is_slow("GET /x") is True
+    # repeated fast samples should pull the EWMA back under the threshold
+    for _ in range(20):
+        p.update("GET /x", 0.01)
+    assert p.is_slow("GET /x") is False
+
+
+def test_predictor_observe_slow_marks_immediately():
+    p = SlowRoutePredictor(threshold=2.0)
+    p.observe_slow("POST /report")
+    assert p.is_slow("POST /report") is True
+
+
+def test_predictor_lru_bound():
+    p = SlowRoutePredictor(threshold=1.0, max_entries=10)
+    for i in range(50):
+        p.update("GET /%d" % i, 0.01)
+    assert len(p._stats) <= 10
+
+
+def test_predictor_empty_key_is_fast():
+    p = SlowRoutePredictor(threshold=1.0)
+    assert p.is_slow(None) is False
+    p.update(None, 5.0)  # must not raise
+    assert p.is_slow(None) is False