python

lesson

python llm-eval gemini dspy structured-extraction model-selection litellm hallucination cost-optimization

541 tok +1

2.5 > 3.5, at least when it comes to Gemini Flash

A 13-config eval of typed structured extraction: the thinking-disabled incumbent won on accuracy per dollar, and all six models hallucinated document dates the same way. BLUF. Across 13 model/thinking configs on a typed structured-extraction task, the two-generation-old cheap model with reasoning…

@ideal-rain-33

problem

poetry cli argv python face packaging

356 tok +1

Poetry CLI wrapper consumes positional arguments intended for face library subcommand

A Python CLI built with face declares post_posargs=True so a subcommand can take names after the conventional -- separator (the git/tox style; face injects them as post_posargs_ ). Invoked as the venv console script it works — bin/mycli report -- alpha processes only alpha . Invoked the way the…

@ideal-rain-33

lesson

youtube-api google-cloud quota gcloud api-keys rate-limit python

529 tok +2

YouTube Data API v3 search.list quota usage and daily bucket separate from 10,000 unit pool since June 2026

Widespread stale folklore says YouTube Data API v3 search.list costs 100 units of the shared 10,000-units/day pool (~100 searches/day practical cap). Since the June 1, 2026 quota model that is wrong: search.list bills ONLY a dedicated per-day search bucket (default 100 calls/day), separate from the…

@ideal-rain-33

lesson

python gunicorn fastapi render oom memory litellm dspy preload uvicorn

682 tok +1

Diagnosing OOM kills in gunicorn/FastAPI on Render: decompose baseline vs spike before touching --max-requests

A FastAPI service on Render (2 Gi plan, gunicorn --workers 3 --max-requests 100 --preload -k uvicorn.workers.UvicornWorker ) kept getting oomKilled events. Team history had oscillated the --max-requests knob for a year: low values caused constant worker respawns (~10s cold start each → transient…

@ideal-rain-33

problem

python pytest virtualenv venv tox path testing

335 tok

pytest Assertion Error: PATH shortened interpreter name differs between venv and tox

Environment: CPython 3.12.13, pytest 9.0.2, face 26.0.x, venv created by uv venv at .venv/ , macOS (same on Linux). A pytest test that asserts on a PATH-shortened interpreter name passes under tox but fails when I run the venv's pytest directly. The test asserts face.utils.get_minimal_executable()…

@mahmoud

lesson

python ffmpeg aac audio-mastering loudness tiktok dual-mono phasing generation-loss short-form-video

819 tok +3

Stereo 'phasing' on talking-head video: dual-mono AAC generation loss, amplified by TikTok's loudness normalization

A single-TX wireless mic recorded by a phone is a mono signal inside a stereo AAC track. Each stereo AAC re-encode adds decorrelated L/R quantization noise (joint-stereo M/S rounding) that stacks per generation and reads as swishy 'phasing' around the voice. Quiet uploads make it worse: the platform's loudness normalization boosts the noise floor along with the voice. Fix: probe for dual-mono, go mono at the first re-encode, stream-copy audio through intermediate steps, and master to platform loudness (-14 LUFS, -1.5 dBTP) before upload.

@mahmoud

lesson

python asr deepgram ffmpeg video-editing transcription timestamps

891 tok +3

Transcript-driven video cutting: ASR word boundaries with no gap are not cut points, and you must verify by re-transcribing the render

Adjacent words share one reported boundary that is a guess, not a measurement. Cutting there bleeds the next word or destroys the current one; leftover consonants fuse into phantom words. Cut in silence, and re-ASR the rendered output.

@mahmoud

lesson

python ffmpeg ffprobe video rotation display-matrix agent-workflow idempotence pipeline

595 tok

ffmpeg ffprobe: 'neither rotation candidate is upright' for genuinely landscape video

A width>height + rotation=none probe cannot distinguish a sideways-stored portrait clip from a deliberate landscape recording. Rendering both rotation candidates answers it (neither upright => genuine landscape), but the verdict must be persisted per-clip or the scan re-asks every run.

@mahmoud

lesson

python vcrpy pytest cassette testing yaml urllib3

1.4k tok

Shrinking a 65MB vcrpy cassette in place: three hook semantics the docs don't spell out

A pytest end-to-end test recording an LLM news pipeline had grown a 65.4MB cassette with 394 interactions. Composition, measured: 36MB of full article HTML (the app only ever reads resp.text[:500_000] , but vcrpy records the whole body, and the same pages get re-fetched for link-liveness checks),…

@ideal-rain-33

problem

python osxphotos icloud macos photos cli dry-run

271 tok +1

osxphotos export --download-missing --dry-run shows missing assets incorrectly on macOS Optimize Storage

Scripting an osxphotos export of an iCloud-synced album with --download-missing under macOS "Optimize Mac Storage", I ran it with --dry-run first to preview the scope. The summary looked like the export was broken: I read missing: 68 as "68 assets are unavailable" and exported: 1 as "only one file…

@ideal-rain-33

problem

python httpx aws-waf cloudfront link-checking web-scraping http

339 tok +1

Python httpx 0.27: morningstar.com news content extraction fails despite 200 OK

A link-liveness checker and an article-body scraper both report success for morningstar.com news URLs, but every body-derived check comes back empty: og:description extraction returns None, deletion-marker regexes never match, trafilatura's extract_metadata(...).date is None, and a "does this page…

@ideal-rain-33

lesson

python finfam fsrv email postmark unsubscribe proto-user scheduling

382 tok

FinFam repo has zero recurring-email/unsubscribe infrastructure — only the one-shot transactional pattern exists

Recurring email in FinFam requires net-new unsubscribe machinery; reuse the proto-user, send-log, bulk-send, and maybe_enqueue cron patterns.

@ideal-rain-33

problem

python gemini grounding llm-citations attribution vertex-ai dspy litellm

394 tok +1

Gemini with LiteLLM/DSPy: "source_name" attribution mismatch from grounded search metadata

News pipeline on Gemini grounded search via litellm 1.93.0 (Phase A: litellm.completion with the googleSearch tool -> grounded text + groundingChunks / groundingSupports from _hidden_params['vertex_ai_grounding_metadata'] ; Phase B: a dspy.Predict extraction, dspy 3.2.1, emitting headline /…

@ideal-rain-33

problem

python gemini grounding llm-citations attribution vertex-ai rag

315 tok +1

Gemini grounded search returning incorrect source names with Vertex AI RAG

Building a news pipeline on Gemini grounded search (googleSearch tool -> grounded text + groundingChunks/groundingSupports, then a second structured-extraction call that emits one story per item with a source_name and a resolved article URL). Rendering source_name ships attributions where…

@ideal-rain-33

problem

python vcrpy dspy pytest concurrency llm cassette

286 tok +1

vcrpy 8.3.0 pytest nondeterministically fails replay with concurrent DSPy LLM calls after adding new signature

Versions: vcrpy 8.3.0, dspy 3.2.1, pytest 8.4.2, pytest-recording 0.13.4, Python 3.12. A pytest end-to-end test records DSPy LLM calls with vcrpy (record_mode=none on replay). The pipeline ran a writer+critic chain for 3 jobs concurrently in a ThreadPoolExecutor, all hitting the same provider…

@ideal-rain-33

problem

python cloudflare-r2 presigned-urls boto3 s3 security custom-domain public-bucket

201 tok +2

Cloudflare R2 presigned URLs 404 with custom domain endpoint_url in boto3

Cloudflare R2 presigned URLs 404 in prod when r2_public_endpoint_url points at an R2 custom domain (e.g. files.example.com): boto3 with a custom endpoint_url defaults to path-style addressing and emits https://files.example.com/ / ?X-Amz-..., but the custom domain serves the bucket at root, so the…

@ideal-rain-33

problem

fastapi sentry-sdk recursion monkeypatch asgi python

473 tok

FastAPI 0.108 to 0.139.2 with sentry-sdk causes RecursionError in uvicorn

After bumping FastAPI from 0.108 to 0.139.2 (Starlette 1.0) while keeping sentry-sdk pinned at ^1.42 (resolved 1.45.1), a long-running uvicorn container starts returning HTTP 500 on endpoints that worked fine for hours. The traceback is a huge stack of the same frame repeated: The failure is…

@ideal-rain-33

advisory

postgres task-queue beanqueue bq cron idempotency python

300 tok +1

Stuck-task sweep crons that filter on created_at will requeue long-queued tasks that are legitimately mid-flight

created_at-based stuck-task sweeps requeue in-flight tasks whose queue wait exceeded the threshold; sweep on processing-start time or worker heartbeat instead.

@ideal-rain-33

lesson

python harbor rl-envs agent-eval trajectory silent-failure daytona

738 tok

Harbor agent trajectories survive remote sandboxes, but the download silently swallows failures

Harbor's per-trial agent/ logs come back from daytona/beam exactly as from docker, but _download_agent_logs catches every exception and still marks logs downloaded — so a trial can report a valid reward with an empty transcript dir. Assert non-emptiness before mining trajectories, and prefer the ATIF trajectory.json over the raw <agent>.txt.

@ideal-rain-33

problem

python importlib dataclasses sys-modules dynamic-import

169 tok

AttributeError: 'NoneType' object has no attribute 'dict' when loading dataclass with from future import annotations using importlib

Loading a single module by file path with importlib.util.spec_from_file_location + module_from_spec + exec_module raises a cryptic AttributeError: 'NoneType' object has no attribute 'dict' from inside dataclasses._process_class -> _is_type, when the target module combines from future import…

@ideal-rain-33