This is the story of how I tracked down four distinct memory leaks across two Python services running on Kubernetes. Each one was a different flavor of “silent accumulation,” and each one taught me something about how easy it is to bleed memory in async Python without any obvious errors.

No names. No product details. Just the bugs, the fixes, and the lessons.

The setup

We had a multi-service platform running on Kubernetes. Two services are relevant here:

• Service A: A document management service. It handled CRUD operations, search indexing, audit logging, and authorization precomputation. Built with FastAPI and async Python, backed by MongoDB and a vector store.

• Service B: A PDF processing service. It parsed documents, extracted text via OCR, and broke content into structured clauses. Heavy on PyMuPDF (fitz) for PDF rendering and Tesseract for OCR.

Both services ran as single-container pods with resource limits. Service A had a 2Gi memory limit. Service B had 4Gi.

Both were getting OOMKilled. Regularly.

The symptoms

Service A’s memory graph looked like this:

Memory (MB)
2048 |                                              ╱ OOMKilled
     |                                           ╱
1536 |                                        ╱
     |                                     ╱
1024 |                                  ╱
     |                               ╱
 512 |                            ╱
     |                         ╱
   0 |________________________╱
     0h        6h        12h        18h        24h

A steady climb at roughly 200MB per hour. No spikes. No correlation with traffic. Just a ramp.

Service B was different - it’d spike during document processing batches and never come back down:

Memory (MB)
4096 |          ╱╲        ╱╲
     |        ╱   ╲     ╱   ╲    ╱ OOMKilled
3072 |      ╱      ╲  ╱      ╲ ╱
     |    ╱         ╲╱        ╱
2048 |  ╱                   ╱
     |╱                   ╱
1024 |                  ╱
     |               ╱
   0 |______________╱
     0h     2h     4h     6h     8h     10h

Every batch pushed memory up. It came down a little, but never to baseline. Classic leak with GC partially reclaiming some objects but not the real offender.

Leak #1: The coroutines that never ran

This was the most insidious one. It looked like perfectly normal Python code:

 1class BaseDoc:
 2    def _write_audit_log(self, action, old_data, new_data):
 3        enqueue_audit_create(self, action, old_data, new_data)
 4
 5    async def save(self, *args, **kwargs):
 6        old = await self.get(self.id)
 7        result = await super().save(*args, **kwargs)
 8        self._write_audit_log("update", old, result)
 9        await create_snapshots(self.id)
10        await insert_versions_docs(self.id)
11        return result

Spot the bug?

_write_audit_log calls enqueue_audit_create(), which is an async function. But _write_audit_log is not async, and it doesn’t await the call. It just calls it.

When you call an async function without await, Python creates a coroutine object but never schedules it. The coroutine sits in memory, holding references to all its arguments - in this case, entire document objects with old_data and new_data.

1Coroutine Object (0x7f...)
2  ├── State: CREATED (never scheduled)
3  └── Frame Object
4       ├── Arg: self (BaseDoc instance, 2KB)
5       ├── Arg: old_data (Dict, 5MB)  <-- LEAKED
6       └── Arg: new_data (Dict, 5MB)  <-- LEAKED
7       └── Local Vars: (captured closure scope)

Python will eventually warn you about this (RuntimeWarning: coroutine was never awaited), but only if you have warnings enabled and you’re looking at the right log. In a noisy production service running dozens of requests per second, these warnings get buried.

The same pattern appeared elsewhere:

 1# In the CRUD endpoints
 2async def create_document(request):
 3    doc = await Document.create(request.data)
 4    enqueue_audit_create(doc)       # ← async, not awaited
 5    qdrant_upserts(doc)             # ← async, not awaited
 6    return doc
 7
 8async def update_document(request):
 9    doc = await Document.update(request.data)
10    enqueue_audit_update(doc)       # ← async, not awaited
11    qdrant_upserts(doc)             # ← async, not awaited
12    return doc

Each unawaited call created a coroutine holding the full document. For a service processing hundreds of documents per hour, that’s hundreds of document-sized objects pinned in memory with no way to be garbage collected.

The fix

Straightforward once you see it:

 1class BaseDoc:
 2    async def _write_audit_log(self, action, old_data, new_data):
 3        await enqueue_audit_create(self, action, old_data, new_data)
 4
 5    async def save(self, *args, **kwargs):
 6        old = await self.get(self.id)
 7        result = await super().save(*args, **kwargs)
 8        await self._write_audit_log("update", old, result)
 9        await create_snapshots(self.id)
10        await insert_versions_docs(self.id)
11        return result

And in the CRUD endpoints:

1async def create_document(request):
2    doc = await Document.create(request.data)
3    await enqueue_audit_create(doc)
4    await qdrant_upserts(doc)
5    return doc

The lesson

Unawaited coroutines are silent memory leaks. They don’t crash. They don’t raise. They just accumulate. And because they hold references to their arguments, they can keep large objects alive long after the request that created them is done.

If you’re running async Python, add this to your linting:

1# In your pytest conftest.py or startup
2import warnings
3warnings.filterwarnings("error", category=RuntimeWarning, message="coroutine.*was never awaited")

This turns the warning into an exception, which makes it impossible to miss.

Leak #2: Creating a new MongoClient on every authorization check

The authorization layer needed to precompute user access contexts. To do this, it queried MongoDB directly (not through the async ODM) using PyMongo’s synchronous MongoClient.

The code looked like this:

 1class AuthorizationPrecompute:
 2    def get_user_permissions(self, user_id):
 3        client = MongoClient(MONGO_URI)
 4        db = client["main_db"]
 5        permissions = list(db.permissions.find({"user": user_id}))
 6        client.close()
 7        return permissions
 8
 9    def get_role_policies(self, role_id):
10        client = MongoClient(MONGO_URI)
11        db = client["main_db"]
12        policies = list(db.policies.find({"role": role_id}))
13        client.close()
14        return policies
15
16    def get_org_settings(self, org_id):
17        client = MongoClient(MONGO_URI)
18        db = client["main_db"]
19        settings = db.settings.find_one({"org": org_id})
20        client.close()
21        return settings
22
23    # ... 10 more methods, each creating its own MongoClient

Thirteen methods. Thirteen MongoClient() instantiations per authorization check.

Each MongoClient creates a connection pool, background SDAM monitoring threads, and internal caches. Even with client.close(), the cleanup isn’t instant - Python’s garbage collector has to reclaim the thread stacks, socket buffers, and internal data structures.

And if an exception happens between MongoClient() and client.close(), the client never gets closed at all. The connection pool lives on, holding threads and sockets until the GC eventually (maybe) collects them.

At 50 requests/second, each triggering an auth check, we were creating 650 MongoClient instances per second. Most were closed, but the transient memory pressure and leaked exceptions meant we were always carrying a few hundred orphaned connection pools.

The fix

One shared client with a connection pool:

 1class AuthorizationPrecompute:
 2    _client = None
 3
 4    @classmethod
 5    def _get_client(cls):
 6        if cls._client is None:
 7            cls._client = MongoClient(
 8                MONGO_URI,
 9                maxPoolSize=20,
10                serverSelectionTimeoutMS=5000
11            )
12        return cls._client
13
14    def get_user_permissions(self, user_id):
15        db = self._get_client()["main_db"]
16        return list(db.permissions.find({"user": user_id}))
17
18    def get_role_policies(self, role_id):
19        db = self._get_client()["main_db"]
20        return list(db.policies.find({"role": role_id}))

Connection overhead went from ~50ms per call (TCP handshake + auth each time) to ~5ms (reused connection from pool).

The lesson

Each MongoClient is an entire connection pool, not a single connection. Creating one per function call is like starting a new database server for every query. Always share a single client instance (or at most one per thread/worker). And if you can’t guarantee close() will be called (i.e., exceptions exist), you definitely need a shared instance.

Leak #3: Cache entries that never expire

The authorization layer also had an in-memory cache for user access contexts:

 1class UserAccessContextCache:
 2    def __init__(self, ttl_seconds=300):
 3        self._cache = {}
 4        self._ttl = ttl_seconds
 5
 6    def get(self, key):
 7        entry = self._cache.get(key)
 8        if entry and time.time() - entry["created"] < self._ttl:
 9            return entry["value"]
10        return None
11
12    def set(self, key, value):
13        self._cache[key] = {
14            "value": value,
15            "created": time.time()
16        }

See the problem?

When get() finds an expired entry, it just returns None. It doesn’t delete the entry. The stale data sits in the dict forever.

And set() only adds entries. Nothing ever removes them.

Over time, this cache grows without bound. Every unique user who hits the service gets a cache entry that is never cleaned up. The access context objects themselves can be large - they contain permission trees, role hierarchies, and org settings.

With 200MB of stale cache entries after a day of operation, this was a significant contributor to the memory ramp.

The fix

Add proactive cleanup:

 1class UserAccessContextCache:
 2    def __init__(self, ttl_seconds=300, cleanup_interval=60):
 3        self._cache = {}
 4        self._ttl = ttl_seconds
 5        self._cleanup_interval = cleanup_interval
 6        self._last_cleanup = time.time()
 7
 8    def _cleanup_expired(self):
 9        now = time.time()
10        if now - self._last_cleanup < self._cleanup_interval:
11            return
12        self._last_cleanup = now
13        expired_keys = [
14            k for k, v in self._cache.items()
15            if now - v["created"] >= self._ttl
16        ]
17        for k in expired_keys:
18            del self._cache[k]
19
20    def get(self, key):
21        self._cleanup_expired()
22        entry = self._cache.get(key)
23        if entry and time.time() - entry["created"] < self._ttl:
24            return entry["value"]
25        # Delete on read if expired
26        if key in self._cache:
27            del self._cache[key]
28        return None
29
30    def set(self, key, value):
31        self._cleanup_expired()
32        self._cache[key] = {
33            "value": value,
34            "created": time.time()
35        }

Cleanup runs at most once per minute and removes all entries older than TTL.

The lesson

Every in-memory cache needs an eviction strategy. If there’s no mechanism to remove old entries, you have a memory leak with extra steps. “TTL on read” is not enough - if users stop requesting a key, that entry stays forever.

Use cachetools.TTLCache or similar if you don’t want to roll your own. Better yet, if the cache is big enough to matter, put it in Redis where you can set EXPIRE and forget about it.

Leak #4: PyMuPDF documents that never closed

Service B was the PDF processing service. It used PyMuPDF (fitz) to render pages and extract text. The endpoints looked like this:

 1@app.post("/parse_pdf")
 2async def parse_pdf(file: UploadFile):
 3    content = await file.read()
 4    doc = fitz.open(stream=content, filetype="pdf")
 5
 6    pages = []
 7    for page_num in range(len(doc)):
 8        page = doc[page_num]
 9        text = page.get_text()
10        pix = page.get_pixmap(dpi=300)
11        image = pix.tobytes("png")
12        pages.append({"text": text, "image": image})
13
14    return {"pages": pages}

No doc.close(). No context manager. No try/finally.

PyMuPDF is a C library wrapper. When you call fitz.open(), it allocates memory in C-land for the document structure, page data, fonts, images, and rendering buffers. Python’s garbage collector can eventually reclaim the Python wrapper object, but the C-side memory is only freed when doc.close() is explicitly called.

While the wrapper theoretically has a __del__ method, relying on it in a high-throughput async loop is dangerous. Python’s GC is lazy and may not run fast enough to keep up with the rate of C-memory allocation, leading to OOMs before the GC even realizes it needs to clean up.

For a 50-page PDF at 300 DPI, the pixmap rendering alone can consume 200–500MB of memory. If an exception occurs mid-processing (corrupt page, unsupported font), the doc handle leaks entirely.

The fix

Two changes: proper resource cleanup, and parallel OCR with controlled concurrency.

 1@app.post("/parse_pdf")
 2async def parse_pdf(file: UploadFile):
 3    content = await file.read()
 4    doc = fitz.open(stream=content, filetype="pdf")
 5    try:
 6        pages = []
 7        for page_num in range(len(doc)):
 8            page = doc[page_num]
 9            text = page.get_text()
10            pix = page.get_pixmap(dpi=300)
11            image = pix.tobytes("png")
12            # Explicitly free the pixmap
13            pix = None
14            pages.append({"text": text, "image": image})
15        return {"pages": pages}
16    finally:
17        doc.close()

This was applied to all four endpoints in the service: /parse_pdf, /process_clauses, /process_clauses_sync, and /extract_links.

We also added parallel strip processing for OCR (controlled by feature flags):

 1import os
 2from concurrent.futures import ThreadPoolExecutor
 3
 4ENABLE_PARALLEL_OCR = os.getenv("ENABLE_PARALLEL_OCR", "false").lower() == "true"
 5STRIP_OCR_WORKERS = int(os.getenv("STRIP_OCR_WORKERS", "8"))
 6
 7def process_page_strips(page, strips):
 8    if not ENABLE_PARALLEL_OCR:
 9        return [ocr_strip(page, s) for s in strips]
10
11    with ThreadPoolExecutor(max_workers=STRIP_OCR_WORKERS) as pool:
12        futures = [pool.submit(ocr_strip, page, s) for s in strips]
13        return [f.result() for f in futures]

The deployment

We didn’t flip everything on at once:

1. Deploy with ENABLE_PARALLEL_OCR=false - just the doc.close() fixes
2. Monitor memory for 24 hours to confirm the leak is fixed
3. Bump resource limits from 4Gi to 6Gi to accommodate parallel processing headroom
4. Enable parallel OCR: ENABLE_PARALLEL_OCR=true

Memory dropped from a steady 4GB to a stable 1.8GB. That’s a 55% reduction from just closing document handles properly.

The lesson

C extension libraries manage their own memory. Python’s GC has no visibility into fitz, numpy, Pillow, or any other C-backed library. If the library provides a close(), use it. If it supports context managers, use with. And always wrap C-heavy processing in try/finally.

The combined impact

Here’s the before and after across both services:

Service A went from daily OOMKills to zero restarts. Service B went from crashing every 8–10 hours to running for weeks without a restart.

How I found them

No fancy tooling. Here’s the actual process:

1. Look at the memory graph. Linear growth = accumulation leak. Sawtooth that doesn’t return to baseline = handle/resource leak.

2. Read the code. Seriously. I read every function in the hot paths and asked “what gets allocated here that doesn’t get freed?”

3. grep for patterns. Once I found one unawaited coroutine, I searched for the pattern across the entire codebase:

1# Find async functions called without await
2rg "^\s+[a-z_]+\(" --type py | grep -v "await " | grep -v "def " | grep -v "#"

Not perfect, but it flagged the 13 MongoClient instantiations immediately.

4. Check warnings. Enabled RuntimeWarning filtering and got 6 unawaited coroutine warnings within the first minute of startup.

5. Kubernetes events. kubectl describe pod shows OOMKilled with the exact memory at death. Compare that to your resource limits and you know how fast it’s growing.

1kubectl get events --field-selector reason=OOMKilling \
2  --sort-by='.lastTimestamp' -n your-namespace

A checklist for your services

If you’re running Python services on Kubernetes and seeing memory growth, check these in order:

• Unawaited coroutines. Search for RuntimeWarning: coroutine.*was never awaited in your logs. Better yet, make it an error.
• MongoClient instantiation. grep -rn "MongoClient(" your_code/ - if it appears in more than one place (especially inside functions, not module level), you probably have a leak.
• In-memory caches without eviction. Any dict that grows with usage and has no del or size cap is a leak. TTLCache, LRU, or maxsize - pick one.
• C extension resources. PyMuPDF, Pillow, lxml, pdfplumber - anything that wraps C needs explicit close() or with statements.
• Large objects in exception handlers. If your except block captures the full traceback and logs the local variables, those locals (which may include entire documents) are pinned until the log entry is flushed.
• Background tasks holding request data. If asyncio.create_task() captures a closure over request objects, those objects live until the task completes.

Final thought

None of these bugs produced an error. No stack trace. No warning (well, one did, but it was buried). The services ran fine - they just got slower and slower until Kubernetes killed them, then they restarted and the cycle began again.

Memory leaks in garbage-collected languages are not the “forgot to call free()” kind. They’re the “I’m holding a reference I didn’t know about” kind. The GC is doing its job perfectly - it’s keeping alive exactly what you told it to.

You just told it wrong.

A coworker said: “I ran task generation, closed the tab, and the status never moved.”

My first reaction was “no way.” The backend is doing the work, not the browser.

Turns out… yeah. It was right. And once you see why, you stop trusting “code after the stream” forever.

This post is about a very specific trap:

• you stream progress with SSE from a FastAPI endpoint
• you put important side effects after the streaming loop (DB status updates, poller kickoff, Slack notifications)
• a client disconnect cancels the stream
• your “after the loop” code never runs

It doesn’t always fail. It fails just enough to make you doubt your own system.

What the system looked like (so you don’t think this is a demo problem)

We had a step-based pipeline. Each step was triggered by an HTTP call from the frontend and did a few things:

1. Receive a request like POST /process-timeline-done (or similar)
2. Start a long-ish worker (in our case an AI agent, but it could be anything)
3. Stream progress back to the UI via Server‑Sent Events
4. Update state in MongoDB (advance a status field)
5. Submit an external job
6. Start a long poller to watch that job (hours sometimes)
7. Send a Slack notification

Only (3) needs a live HTTP connection. The rest should be able to finish even if the user refreshes or closes the tab.

We had code split into things like:

The bug was that our implementation accidentally made a bunch of “No” things depend on HTTP anyway.

The bug (the exact pattern)

This is the shape of the broken code:

1# ❌ Broken: critical work is after streaming.
2async def process_step():
3    async for update in agent():
4        yield sse(update)  # disconnect can cancel/raise here
5
6    # If the client disconnects, this may never run.
7    await db.update_status("tasks_done")
8    asyncio.create_task(start_polling(job_id))
9    await slack.notify("done")

It feels reasonable because it matches how we think:

“Stream updates while running… then when the loop ends, do the completion stuff.”

But in a streaming response, the loop ending is not guaranteed. The client can just leave.

What actually happens when the client disconnects

Let’s make it concrete with a timeline.

Normal happy path:

Browser                  FastAPI/Starlette           MongoDB / External
   |                           |                          |
   | POST /process-step        |                          |
   |-------------------------->|                          |
   | 200 OK (SSE stream)       |                          |
   |<--------------------------|                          |
   | data: update #1           |                          |
   |<--------------------------|                          |
   | data: update #2           |                          |
   |<--------------------------|                          |
   | ...                       |                          |
   | data: final               |                          |
   |<--------------------------|                          |
   | stream ends               |                          |
   |                           | update status            |
   |                           |------------------------->|
   |                           | start poller             |
   |                           |------------------------->|
   |                           | Slack notify             |
   |                           |------------------------->|

Disconnect path:

Browser                  FastAPI/Starlette           MongoDB / External
   |                           |                          |
   | POST /process-step        |                          |
   |-------------------------->|                          |
   | 200 OK (SSE stream)       |                          |
   |<--------------------------|                          |
   | data: update #1           |                          |
   |<--------------------------|                          |
   | data: update #2           |                          |
   |<--------------------------|                          |
   | [tab closed]              |                          |
   X                           |                          |
                               | stream task canceled     |
                               | (right on a yield/send)  |
                               |                          |
                               | ❌ status update skipped  |
                               | ❌ poller never started   |
                               | ❌ Slack never sent       |

The punchline is simple: SSE streaming is not “just a loop.” It’s a loop being consumed by a response writer that is allowed to stop at any time.

Under the hood (why Starlette cancels your generator)

This is not the exact Starlette source code, but it matches the design.

Streaming responses typically run two things concurrently:

• a task that iterates your generator and sends chunks (http.response.body)
• a task that listens for http.disconnect

When the disconnect happens, the server cancels the streaming task. That cancellation can hit while the framework is trying to send your last chunk… which was produced by your generator… which means your generator gets torn down mid-flight.

Two practical details that matter a lot:

1. Cancellation shows up as asyncio.CancelledError (or sometimes as “write failed” type errors depending on stack).
2. In modern Python, asyncio.CancelledError is a BaseException, not an Exception. So except Exception: won’t catch it.

That second one is why “but I wrapped it in try/except” doesn’t save you.

A minimal repro you can run in two minutes

Here’s a tiny app that demonstrates the exact failure mode. The print() after the loop is your “update MongoDB status” in disguise.

 1import asyncio
 2from fastapi import FastAPI
 3from starlette.responses import StreamingResponse
 4
 5app = FastAPI()
 6
 7async def gen():
 8    for i in range(1000):
 9        yield f"data: {i}\n\n".encode()
10        await asyncio.sleep(0.2)
11    print("stream ended; now do side effects")
12
13@app.get("/sse")
14async def sse():
15    return StreamingResponse(gen(), media_type="text/event-stream")

Run it, then:

1curl -N http://localhost:8000/sse

Wait for a few numbers, hit ctrl+c in the curl terminal.

You will not see the “now do side effects” line.

That’s the bug.

The real fix: stop making your job depend on the HTTP stream

Once you accept “the client can vanish at any time”, the design gets boring (in a good way):

• The job runs to completion in a background task.
• Streaming is just a best-effort view of progress.

I like to think in terms of ownership:

• The worker owns side effects (DB writes, poller, Slack).
• The request handler owns only streaming bytes to the client.

What I tried first (and why it didn’t help): BackgroundTasks

FastAPI has BackgroundTasks. It’s useful, but it’s not a magic “disconnect-proof” button.

The catch is simple: background tasks are executed after the response is done.

For a normal JSON response, “response is done” means “we returned a body, the server wrote it, done.”

For an SSE response, “response is done” means “the stream ended.”

Which is… exactly the thing we can’t rely on.

If the client disconnects and the streaming response gets canceled, your “run this after the response” work is still hanging off the same lifecycle. It’s not the separation you actually want.

“Can’t I just move the status update into finally?”

This is another tempting fix:

1async def gen():
2    try:
3        async for update in agent():
4            yield sse(update)
5    finally:
6        await db.update_status("tasks_done")

Sometimes it works. Sometimes it doesn’t. It depends on where cancellation lands.

Cancellation can interrupt awaits inside finally too. So you end up with a fun situation where you thought you made it reliable, but you just changed the probability of it failing.

If you really want to run cleanup under cancellation, you need to start thinking about shielding and cancel scopes. At that point you’re already halfway to the “worker owns completion” design anyway, so I’d rather just do it properly.

The queue pattern (with one important correction)

The common implementation is “worker pushes updates into a queue, streamer drains it.”

That’s fine, but there’s a nasty gotcha:

If the client disconnects and the streamer stops draining, a bounded queue can block your worker forever.

So don’t do this in your worker:

1await queue.put(update)  # ❌ can hang if consumer is gone

If you do, you just recreated the original bug: completion is again tied to streaming, just indirectly.

What I do instead:

• Keep the queue bounded (to protect memory)
• Put updates in a non-blocking way and drop when full
• Stop queueing progress once the client is gone

Here’s a practical sketch:

 1import asyncio
 2import json
 3import logging
 4
 5logger = logging.getLogger(__name__)
 6
 7def sse(data: dict) -> bytes:
 8    return f"data: {json.dumps(data)}\n\n".encode("utf-8")
 9
10def try_put(queue: asyncio.Queue, item) -> None:
11    try:
12        queue.put_nowait(item)
13    except asyncio.QueueFull:
14        # progress is optional; completion isn't
15        pass
16
17def log_task_result(task: asyncio.Task) -> None:
18    try:
19        task.result()
20    except asyncio.CancelledError:
21        pass
22    except Exception:
23        logger.exception("background worker crashed")
24
25async def run_to_completion(queue: asyncio.Queue, client_gone: asyncio.Event):
26    job_id = None
27    try:
28        async for update in agent():
29            job_id = job_id or update.get("job_id")
30            if not client_gone.is_set():
31                try_put(queue, update)
32
33        # ✅ These must run even if client disconnects.
34        await db.update_status("tasks_done")
35        asyncio.create_task(start_polling(job_id))
36        await slack.notify("done")
37    except Exception as e:
38        if not client_gone.is_set():
39            try_put(queue, {"final": True, "success": False, "error": str(e)})
40        raise
41    finally:
42        if not client_gone.is_set():
43            try_put(queue, None)  # sentinel for streamer
44
45async def stream(queue: asyncio.Queue, client_gone: asyncio.Event):
46    try:
47        while True:
48            item = await queue.get()
49            if item is None:
50                break
51            yield sse(item)
52    except asyncio.CancelledError:
53        # disconnected: stop streaming, signal worker to stop enqueueing
54        client_gone.set()
55        raise
56    finally:
57        client_gone.set()
58
59async def handler():
60    queue = asyncio.Queue(maxsize=200)
61    client_gone = asyncio.Event()
62
63    task = asyncio.create_task(run_to_completion(queue, client_gone))
64    task.add_done_callback(log_task_result)
65
66    return StreamingResponse(stream(queue, client_gone), media_type="text/event-stream")

Is it perfect? No. But it has the right shape:

• worker can finish without a consumer
• progress streaming is best-effort
• failures are logged instead of silently disappearing

If you want something more durable (client reconnects and can replay progress), you need to store progress somewhere real (Redis/pubsub, DB, etc.). In-memory queues won’t do that. But that’s a different problem.

The architecture I like better (when this gets serious)

The queue pattern is fine when:

• you mostly care about “don’t lose completion when the client disconnects”
• you don’t care about reconnecting and replaying a progress history

But once you ship this to real users, you’ll hit two common problems:

1. Browsers reconnect SSE automatically (especially EventSource), and you don’t want a reconnect to re-run the job.
2. People refresh and expect to see progress again (or at least not lose the final result).

So the more robust shape is:

1. POST /steps/{step}/start → returns a run_id
2. Worker runs independent of any SSE connection and persists state (status + maybe progress)
3. GET /runs/{run_id}/events → SSE stream that is purely “subscribe”, not “start”

Now reconnecting is safe. You might lose some live updates, but you can always read the current status from MongoDB, and you can keep polling running even if nobody is watching.

If you want a really clean UX, you can also persist event ids and support Last-Event-ID, but that’s optional. The big win is separating “start work” from “watch work.”

SSE details that are boring but matter

SSE is “simple” in the same way DNS is “simple.”

It works great until you put it behind real infrastructure.

Event formatting is strict

Each event ends with a blank line. If you forget the blank line, the browser buffers and you’ll think streaming is broken.

Simplest possible event:

1data: {"message":"hi"}

(Yes, that extra newline is part of it.)

Keepalives

If your worker can go quiet for a while, send a comment every ~15–30 seconds so proxies don’t decide the connection is “idle”:

1yield b": keep-alive\n\n"

Nginx buffering

If your stream arrives in one big chunk at the end, it’s usually Nginx buffering.

You don’t fix buffering with Python. You fix it with config.

At minimum, you want buffering off for the SSE location and timeouts that won’t murder long streams.

Something like this (not gospel, but you get the idea):

1location /runs/ {
2  proxy_http_version 1.1;
3  proxy_set_header Connection "";
4  proxy_buffering off;
5  proxy_cache off;
6  proxy_read_timeout 3600s;
7  add_header X-Accel-Buffering no;
8}

The other gotchas (the stuff that actually bites later)

Don’t use request-scoped dependencies in your background worker

If you have Depends(get_db) where get_db is a yield dependency, it gets cleaned up when the request is done.

If your worker is still running, it may now be holding a dead session/client.

Solution: the worker should acquire its own long-lived clients (Mongo, HTTP client, etc.) that are not tied to the request lifecycle.

Treat status updates like a state machine

If you do timeline_done → tasks_done → final_done, don’t just write tasks_done unconditionally.

Do “only update if current status is timeline_done” (compare-and-set). Otherwise retries and races will eventually mess up your state.

Don’t confuse “survives client disconnect” with “durable”

This pattern survives a client disconnect. It does not survive:

• the process dying
• the pod being evicted
• a deploy that SIGTERMs the worker mid-step

If the job truly must finish no matter what, use a real job system. SSE should subscribe to a job id, not be the job.

Decide if disconnect should cancel the job (explicitly)

Some people will ask for “closing the tab cancels the job.”

That’s a separate feature. Don’t accidentally implement it via “oops, the client disconnected so the worker got canceled.”

If you want cancellation, do it explicitly:

• create a job id
• store cancel intent
• have the worker check it and stop cleanly

Client disconnect is not a reliable cancel signal. It’s just the network being the network.

A boring checklist (the stuff I verify before I trust the fix)

• Start a step, then close the tab mid-stream → status still advances later
• Start a step, refresh rapidly → no duplicate runs / no broken state transitions
• Simulate a proxy timeout → completion still happens
• Kill the SSE client (curl -N ... then ctrl+c) → completion still happens
• Worker crashes mid-way → status reflects failure, and it doesn’t get stuck “in progress” forever

Performance notes (because someone will ask)

The queue approach is usually cheap, but it’s not free.

• Memory: queue_size * avg_event_size per in-flight stream. Keep event payloads small.
• CPU: queue ops are basically noise; the expensive bits are your DB/network calls.
• DB load: if you add read-after-write verification (I sometimes do), it’s extra reads. Worth it if your system lies to you otherwise.

The bigger risk isn’t CPU. It’s “we launched 500 background tasks because 500 people clicked the button.” Put a cap somewhere.

Deployment strategy (how I roll it out without getting yelled at)

1. Staging: disconnect tests, refresh tests, and “kill the client” tests.
2. Canary: small percentage of traffic. Watch for stuck statuses.
3. Full rollout: watch completion rate for a day. If it drops, roll back fast.

How I tested it (and how I’d recommend you test it)

Nothing fancy. I just did the thing that used to break it:

• start the step
• wait for a couple SSE updates
• close the tab / refresh / kill the connection
• watch MongoDB and server logs

What I wanted to see is boring:

• status advances anyway
• poller starts anyway
• Slack fires anyway

Once you watch the system finish without a client attached, you stop trusting “after the stream” code forever.

And honestly you should. It’s a trap.