본문으로 건너뛰기
cccskills
무료GitHub에 공개

asyncopenai-concurrency-httpx-pool

Raise real concurrency in asyncio LLM batch scorers built on the OpenAI SDK (AsyncOpenAI, including OpenAI-compatible providers like DeepSeek). Use when: (1) raising an asyncio.Semaphore above ~100 produces no throughput gain, (2) a batch pipeline saturates near 100 in-flight requests despite a larger semaphore, (3) planning a high-concurrency campaign against a provider with no hard rate limit (DeepSeek v4-flash tolerates 2000+ in flight). Root cause: AsyncOpenAI's default httpx pool caps max_connections at 100, silently bottlenecking any larger semaphore — you must pass a custom http_client with httpx.Limits sized to the semaphore.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

kennethkhoocy/applied-micro-skills
연동 도구OpenAI
설치 방법 보기

포함된 파일(2)

  • SKILL.md3.9 KB
  • README.md3.5 KB

SKILL.md(원문)

설치하기 전에 에이전트에게 전달되는 지시 내용을 확인할 수 있습니다.

AsyncOpenAI Concurrency: the Hidden httpx Pool Cap

Problem

Async batch scorers typically gate concurrency with asyncio.Semaphore(N). Raising N above ~100 silently does nothing: the OpenAI SDK's default httpx transport caps the connection pool at max_connections=100, so excess tasks queue inside httpx instead of reaching the provider. The semaphore looks like the throttle but is not the binding constraint — there is no error, just a throughput ceiling.

Context / Trigger Conditions

  • asyncio.Semaphore(N) with N > 100 around client.chat.completions.create shows the same throughput as N = 100
  • Client constructed as AsyncOpenAI(api_key=..., base_url=...) with no http_client argument (the default transport)
  • Provider is known to allow high concurrency (DeepSeek v4-flash: ~2500)
  • Symptom check: requests-in-flight measured at the server never exceeds ~100

Solution

Size the httpx pool to the semaphore when constructing the client:

import httpx
from openai import AsyncOpenAI

CONCURRENCY = 2000
client = AsyncOpenAI(
    api_key=..., base_url="https://api.deepseek.com",
    http_client=httpx.AsyncClient(limits=httpx.Limits(
        max_connections=CONCURRENCY,
        max_keepalive_connections=CONCURRENCY)))
sem = asyncio.Semaphore(CONCURRENCY)

Both edits are required; either alone caps the other. Keep the per-request retry loop — at high concurrency transient failures are more likely, and the retry envelope is what turns them into non-events.

Verification

Throughput scales with N. Verified 2026-07-17 on DeepSeek v4-flash (deepseek-chat, JSON-mode unit scoring, ~1.5k-token prompts): at semaphore 50 a cold 13.8k-request chunk took ~70 min; at semaphore 2000 + matched pool, a 14.4k-request chunk took ~11 min (~40 req/s sustained, ~250/s burst on a 1.7k-request tail chunk), 0 failed requests, 0 schema-invalid responses. Effective speedup ~7x rather than 40x — server-side queuing absorbs the rest — but with zero reliability cost.

Example

Specialist Directors US T1 campaign: final 8 chunks (100,243 calls) completed in ~55 minutes for $8.08 after the fix, versus a projected ~9 hours at the old setting. The edit is two lines in the scorer; the semaphore constant alone would have been a silent no-op.

Notes

  • DeepSeek publishes no hard rate limit and handled 2000 in-flight cleanly; the practical ceiling reported is ~2500. Other providers enforce RPM/TPM caps — check before sizing.
  • Windows: default asyncio proactor loop handled 2000 sockets without tuning; no ulimit-style adjustment needed.
  • Companion ops lesson from the same campaign: when a run's plan changes scale (two-night legs -> one-shot), re-audit the launch glue's hardcoded limits — a wrapper MAX-HOURS=10 safety cap sized for the old plan hard-killed a healthy runner at 97/108 chunks. Caps and budgets in supervisor scripts must be revisited whenever expected duration changes.
  • Concurrency is a pure throughput knob: per-request outputs are unchanged (temperature 0, independent requests), so raising it mid-campaign does not create a scoring seam — unlike model/prompt changes, which do (see [llm-campaign-drift-gate]).

리뷰

아직 리뷰가 없습니다. 사용해 본 소감을 남겨 주세요.

댓글

로그인 없이 쓸 수 있습니다. 사용 후기, 질문, 작동하지 않은 환경 등을 편하게 남겨 주세요.

아직 댓글이 없습니다. 첫 댓글을 남겨 보세요.

보내면 바로 공개됩니다. 이메일 주소나 전화번호 같은 개인 정보는 쓰지 마세요. URL은 2개까지입니다.

0 / 1,000자

이 양식은 reCAPTCHA로 보호되며 Google 개인정보처리방침 및 서비스 약관이 적용됩니다.

같은 저장소의 스킬

adjudication-sheets

무료

Build human adjudication / hand-labeling sheets from LLM-pipeline data without evidence truncation. Use when: (1) preparing a CSV/Excel sheet for a human to rule on cases an LLM classifier or rater panel judged, (2) a labeler reports "there is no information to label from" or cells look empty in Excel, (3) excerpt columns cluster at one exact length (e.g. all 1,500 chars — a hard truncation cap). Covers: full rating-basis recovery, Excel 32,767-char cell cap, multi-line CSV mangling, ruling dropdowns, companion text files.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

kennethkhoocy/applied-micro-skills222026년 9월 5일 업데이트

adversarial-empirical-review

무료

N-round adversarial review pipeline for empirical research output — the chain from data to LaTeX tables to a manuscript that cites them. A Claude drafter proposes minimal diffs, a deterministic mechanical battery gates every diff from a clean state with a regression gate, a Codex reviewer files check-backed critiques, and a blind judge panel decides residual disputes. Manual-invoke ONLY: trigger when the user explicitly runs /adversarial-empirical-review or names 'adversarial-empirical-review' / 'adversarial empirical review'. Do NOT auto-trigger on generic 'review my results', 'check my tables', or manuscript-editing requests. For prose-style refinement use style-emulation instead; this skill AUDITS WHETHER THE TABLES ARE CORRECT — that each number in the tables is what the analysis code computes, reproduces from the data, and is internally consistent. It is an empirical + code review: the manuscript is read only to resolve table numbering, and prose is not examined.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

kennethkhoocy/applied-micro-skills222026년 9월 5일 업데이트

annotator-input-parity-check

무료

Before designing, training, or auditing ANY model that replicates human-annotated labels, audit the annotation protocol's INPUT — the exact document/evidence the human labelers consulted — and give the model that same input. Use when: (1) designing a classifier/LLM extractor whose target is a hand-coded label set, (2) a label-replication model shows low recall concentrated in a label subset and the diagnosis on offer is "the label's information is not in the features", (3) reviewers propose construct splits (e.g. "designation vs record-evident"), adjudication sittings, or per-domain stop rules to explain residual disagreement with gold, (4) validating an extraction pipeline against labels transcribed from a source document. Symptom of the underlying failure: elaborate theory accumulates to explain why gold is "partially unpredictable" when the model was simply never shown the document the annotators read.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

kennethkhoocy/applied-micro-skills222026년 9월 5일 업데이트

cite-placement

무료

Place pre-screened literature citations into a LaTeX or Word manuscript, or restyle the citations already in one. Three modes: (1) inline placement — inline \cite{}/\citet{}/\citep{} with a compiled references.bib, for author-date journals (APA, MLA, Harvard, Chicago author-date, IEEE, Vancouver); (2) footnote placement — full formatted \footnote{} or OOXML footnotes for legal and notes styles (Bluebook, OSCOLA, Chicago, APA, McGill) with Id./supra short forms; (3) restyle — convert existing footnote citations from one style to another. This skill is manual-invoke ONLY — trigger ONLY when the user explicitly runs /cite-placement or explicitly names the "cite-placement" skill. Do NOT auto-trigger on general citation, footnote, or reference requests.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

kennethkhoocy/applied-micro-skills222026년 9월 5일 업데이트

download-gated-pdfs

무료

Download the actual PDF binary from bot-gated sites (taxpolicycenter.org, urban.org, SSRN-hosted mirrors, think-tank/publisher sites) via the Wayback Machine id_ URL form. Use when: (1) curl/WebFetch of a .pdf URL returns HTML instead of a PDF even with a browser User-Agent, (2) pypdf fails with "invalid pdf header: b'<!DOC'" or "EOF marker not found" on a freshly downloaded file, (3) Firecrawl can parse the PDF to markdown but you need the original file on disk (e.g., filing a reference copy).

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

kennethkhoocy/applied-micro-skills222026년 9월 5일 업데이트

event-study-cars

무료

Complete methodology for computing publication-quality cumulative abnormal returns with proper event-study test statistics, matching the robustness of Kaspereit's eventstudy2 for Stata. Covers dateline construction, event-date mapping, estimation and event windows, thin-trading adjustment, OLS with Theil prediction error correction, abnormal return computation, CAR/CAAR/AAR accumulation, boundary contamination guards, and common tests such as Patell, BMP, Kolari-Pynnonen, generalized sign, Wilcoxon, and GRANK-T. Use when the user mentions abnormal returns, event windows, market-model regressions, CARs, CAAR, AAR, eventstudy2, thin trading, trade-to-trade returns, or event-study test statistics.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

kennethkhoocy/applied-micro-skills222026년 9월 5일 업데이트

kennethkhoocy의 스킬 모두 보기

이 스킬의 문제 신고