audio streaming app
plyr.fm
fix: migrate rate limiting from in-memory to Redis (#1044)
rate limits were tracked per Fly Machine instance, so 2 machines meant
2x the configured limit. use the existing docket Redis for global
counters. falls back to memory:// when DOCKET_URL is not set (local dev).
closes #1043
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
authored by
Claude Opus 4.6
and committed by
GitHub
67ebb28f
5153d514
+1
-1
backend/src/backend/utilities/rate_limit.py
+1
-1
backend/src/backend/utilities/rate_limit.py
+35
backend/tests/test_rate_limit.py
+35
backend/tests/test_rate_limit.py
···
1
+
"""Tests for rate limiting configuration."""
2
+
3
+
from unittest.mock import patch
4
+
5
+
from backend.config import settings
6
+
7
+
8
+
def test_limiter_uses_redis_when_docket_url_set() -> None:
9
+
"""limiter should use docket Redis URL for storage when available."""
10
+
with patch.object(settings.docket, "url", "redis://localhost:6379/0"):
11
+
# re-import to pick up patched settings
12
+
import importlib
13
+
14
+
import backend.utilities.rate_limit as rl_module
15
+
16
+
importlib.reload(rl_module)
17
+
18
+
assert rl_module.limiter._storage_uri == "redis://localhost:6379/0"
19
+
20
+
# reload again to restore original state
21
+
importlib.reload(rl_module)
22
+
23
+
24
+
def test_limiter_falls_back_to_memory_when_no_docket_url() -> None:
25
+
"""limiter should fall back to in-memory storage when DOCKET_URL is empty."""
26
+
with patch.object(settings.docket, "url", ""):
27
+
import importlib
28
+
29
+
import backend.utilities.rate_limit as rl_module
30
+
31
+
importlib.reload(rl_module)
32
+
33
+
assert rl_module.limiter._storage_uri == "memory://"
34
+
35
+
importlib.reload(rl_module)
+5
-12
docs-internal/rate-limiting.md
+5
-12
docs-internal/rate-limiting.md
···
13
13
| `RATE_LIMIT_ENABLED` | `true` | Enable/disable rate limiting globally. |
14
14
| `RATE_LIMIT_DEFAULT_LIMIT` | `100/minute` | Global limit applied to all endpoints by default. |
15
15
| `RATE_LIMIT_AUTH_LIMIT` | `10/minute` | Strict limit for auth endpoints (`/auth/start`, `/auth/exchange`). |
16
-
| `RATE_LIMIT_UPLOAD_LIMIT` | `5/minute` | Strict limit for file uploads (`/tracks/`). |
16
+
| `RATE_LIMIT_UPLOAD_LIMIT` | `20/minute` | Strict limit for file uploads (`/tracks/`). |
17
17
18
18
## Architecture
19
19
20
-
The current implementation uses **in-memory storage**.
20
+
The implementation uses **Redis-backed storage** via the existing docket Redis instance (`DOCKET_URL`).
21
21
22
-
* **Per-Instance:** Limits are tracked per application instance (Fly Machine).
23
-
* **Scaling:** With multiple replicas (e.g., 2 machines), the **effective global limit** scales linearly.
24
-
* Example: A limit of `100/minute` with 2 machines results in a total capacity of roughly `200/minute`.
22
+
* **Global Counters:** Limits are shared across all application instances (Fly Machines). A `100/minute` limit means 100 requests total, regardless of which machine handles them.
25
23
* **Keying:** Limits are applied by **IP address** (`get_remote_address`).
26
-
27
-
### Why in-memory?
28
-
For our current scale, in-memory is sufficient and avoids the complexity/cost of a dedicated Redis cluster. This provides effective protection against single-source flooding (DDoS/brute-force) directed at any specific instance.
29
-
30
-
### Future State (Redis)
31
-
If strict global synchronization or complex tier-based limiting is required in the future, we will migrate to a Redis-backed limiter. `slowapi` supports Redis out of the box, which would allow maintaining shared counters across all application instances.
24
+
* **Fallback:** When `DOCKET_URL` is not set (e.g., local dev without Redis), falls back to in-memory storage automatically.
32
25
33
26
## Adding Limits to Endpoints
34
27
···
50
43
51
44
## Monitoring
52
45
53
-
Rate limit hits return `429 Too Many Requests`. These events are logged and will appear in Logfire traces with the `429` status code.
46
+
Rate limit hits return `429 Too Many Requests`. These events are logged and will appear in Logfire traces with the `429` status code.