Make configurable via environment variable & fix typos in profiler trace filenames

[iamfaham]

Make profiling_directory configurable via environment variable & fix typos in profiler trace filenames

Search Before Asking

• I searched the Issues and did not find a duplicate.
• Related: Workflows fail to stop while workflow inference profiling is enabled (readonly filesystem) #1110 (read-only filesystem profiling crash, fixed by PR fix: handle OSError on read-only filesystem during profiling #2732)

---

Feature Description

Following the merge of PR #2732 (which addressed the read-only filesystem crash from #1110), the profiling subsystem now gracefully handles OSError when writing traces. However, the output directory itself is still hardcoded as "./inference_profiling" in two separate locations with no environment variable override — making it the only pipeline/profiling setting that can't be configured at deploy time.

Current state

In inference/core/env.py (lines 1036–1037), the profiling feature already has two env vars:

ENABLE_WORKFLOWS_PROFILING = str2bool(os.getenv("ENABLE_WORKFLOWS_PROFILING", "False"))
WORKFLOWS_PROFILER_BUFFER_SIZE = int(os.getenv("WORKFLOWS_PROFILER_BUFFER_SIZE", "64"))

And the sibling pipeline configs all follow the env-var pattern:

Setting	Env Var	Default
Predictions queue size	INFERENCE_PIPELINE_PREDICTIONS_QUEUE_SIZE	512
Restart attempt delay	INFERENCE_PIPELINE_RESTART_ATTEMPT_DELAY	1
Video source buffer size	VIDEO_SOURCE_BUFFER_SIZE	64
Profiling enabled	ENABLE_WORKFLOWS_PROFILING	False
Profiler buffer size	WORKFLOWS_PROFILER_BUFFER_SIZE	64
Profiling directory	❌ none	hardcoded "./inference_profiling"

The hardcoded default appears in:

1. inference/core/interfaces/stream/inference_pipeline.py line 503 — init() method parameter default
2. inference_sdk/http/entities.py line 150 — InferenceConfiguration dataclass field default

Proposed change

Add a new environment variable WORKFLOWS_PROFILING_DIRECTORY to inference/core/env.py:

WORKFLOWS_PROFILING_DIRECTORY = os.getenv("WORKFLOWS_PROFILING_DIRECTORY", "./inference_profiling")

Then import and use it as the default in both locations above instead of the hardcoded string.

Why this matters

1. Containerised deployments (Docker, K8s): The working directory inside a container is often read-only or ephemeral. While PR fix: handle OSError on read-only filesystem during profiling #2732 made the crash non-fatal, users still can't direct profiling output to a writable mounted volume without modifying the source code or passing the directory through the Python API every time.

2. Consistency: Every other profiling/pipeline knob has an env var. The directory is the sole exception — an easy oversight to fix.

3. SDK usage: InferenceConfiguration in the SDK also hardcodes the same path. An env var gives a single, code-free control point for both the HTTP server and SDK clients.

4. Observability pipelines: In production, teams often want traces written to a shared volume (e.g. /data/profiling) for collection by a sidecar. An env var makes this a one-line Docker Compose / Helm values change.

---

Bonus: Typo fixes in profiler trace filenames

Two typos exist in the profiler trace output that produce malformed filenames:

1. Missing 'r' in track — inference/core/interfaces/stream/utils.py line 142

# Current (typo):
f"inference_pipeline_workflow_execution_tack_{formatted_time}.json"
# Should be:
f"inference_pipeline_workflow_execution_track_{formatted_time}.json"

2. Same typo in the SDK — inference_sdk/http/utils/profilling.py line 21

# Current (typo):
f"workflow_execution_tack_{formatted_time}.json"
# Should be:
f"workflow_execution_track_{formatted_time}.json"

3. missmatch → mismatch — inference/core/interfaces/stream/utils.py line 43

# Current (typo):
error_description="Cannot apply `video_source_properties` to video sources due to missmatch in "
# Should be:
error_description="Cannot apply `video_source_properties` to video sources due to mismatch in "

Note: missmatch appears in several other files across the codebase (stream_manager, workflows, inference_models). Those could be addressed in a follow-up cleanup PR if desired, but the two above are directly in the profiling code path.

---

Example Usage

Docker

docker run \
  -e ENABLE_WORKFLOWS_PROFILING=True \
  -e WORKFLOWS_PROFILING_DIRECTORY=/data/profiling \
  -v /host/profiling:/data/profiling \
  roboflow/inference:latest

Docker Compose

services:
  inference:
    image: roboflow/inference:latest
    environment:
      ENABLE_WORKFLOWS_PROFILING: "True"
      WORKFLOWS_PROFILING_DIRECTORY: "/data/profiling"
    volumes:
      - ./profiling:/data/profiling

Helm / K8s

env:
  - name: ENABLE_WORKFLOWS_PROFILING
    value: "True"
  - name: WORKFLOWS_PROFILING_DIRECTORY
    value: "/mnt/profiling"

---

Suggested implementation (happy to submit a PR)

1. inference/core/env.py — add after line 1037:

   WORKFLOWS_PROFILING_DIRECTORY = os.getenv("WORKFLOWS_PROFILING_DIRECTORY", "./inference_profiling")

2. inference/core/interfaces/stream/inference_pipeline.py line 503 — change default:

   profiling_directory: str = WORKFLOWS_PROFILING_DIRECTORY,

3. inference_sdk/http/entities.py line 150 — change default:

   profiling_directory: str = WORKFLOWS_PROFILING_DIRECTORY,

4. Fix typos in utils.py (lines 43, 142) and profilling.py (line 21) as described above.

[iamfaham]

Happy to submit a PR for this if the direction looks right.