From 57744c532c58d3e2bd248b4368ecba0a42402ed8 Mon Sep 17 00:00:00 2001 From: Artyom Date: Thu, 5 Nov 2026 23:59:54 +0300 Subject: [PATCH] Cache the immutable message statement and validate scaling evidence --- AGENTS.md | 4 +- .../changes/remeasure-lab3-scaling/design.md | 14 ++++ .../changes/remeasure-lab3-scaling/tasks.md | 6 +- scripts/scaling.py | 3 + .../messaging/infrastructure/postgres.py | 3 +- student-tests/api/test_idempotency.py | 20 +++-- tools/plot_track1.py | 76 ++++++++++++------- 7 files changed, 87 insertions(+), 39 deletions(-) diff --git a/AGENTS.md b/AGENTS.md index 035ec38d..440c503f 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -27,9 +27,9 @@ make limits # уровень 4: лимиты, порты, с make detection # уровень 4: время обнаружения SIGTERM/kill/pause/PostgreSQL против заявленного максимума make failover-each # уровень 4: каждая реплика по очереди под k6 (RATE=100) → evidence/failover-each make rolling-update [FROM_REV=<коммит>] # уровень 4: замена реплик по одной под k6, migrate один раз -make track1 # трек 1: снимок хоста, серии A (1x4/2x2/4x1) и B (1x1/2x1/8x1), до/после на 8x1, контроль, графики (≈3,5 ч) +make track1 # трек 1: снимок хоста, серии A (1x4/2x2/4x1) и B (1x1/2x1/8x1), до/после на 8x1, контроль, графики; volumes сохраняются make track1-control # трек 1: пул 4/8/16 и maxconn 16/32/64 на 8x1; 8x0.5 против 4x1 на одних CPU -make plots # трек 1: графики и сводка потолков из evidence/track1/*/summary.csv +make plots # трек 1: графики и сводка потолков из evidence/track1-sqlalchemy/*/summary.csv make down # остановка без удаления данных (с профилем bench); make reset-data — удаление volumes ``` diff --git a/openspec/changes/remeasure-lab3-scaling/design.md b/openspec/changes/remeasure-lab3-scaling/design.md index 3244377b..a8dcd62d 100644 --- a/openspec/changes/remeasure-lab3-scaling/design.md +++ b/openspec/changes/remeasure-lab3-scaling/design.md @@ -60,3 +60,17 @@ Context7: /sqlalchemy/sqlalchemy/rel_2_0_45, core/connections.rst, SQL compilati caching и lambda_stmt. Прямой lambda_stmt(post_message) отвергнут библиотекой из-за timedelta внутри функции; enable_tracking=False допустим для полностью статического дерева с явными bindparam. Версия проверена локально: 2.0.45. + +Для графиков используется Matplotlib 3.11.2 (версия закреплена в make plots), backend Agg. +Context7 подтвердил PNG-рендеринг; поле Description для даты/SHA дополнительно сверено с +[официальным API PNG](https://matplotlib.org/stable/api/backend_agg_api.html#matplotlib.backends.backend_agg.FigureCanvasAgg.print_png). + +## Наблюдения базовых серий + +A: 54 прогона; пики 2687/2532/2422 rps для 1×4/2×2/4×1. +B: 36 прогонов; 8×1 даёт около 3420 rps при 3600 предложенных, дальше +пик не растёт. На 5400 медианы: около 32 соединений ждут Lock, 16 idle in transaction +из пула 64; PG CPU около 158/200%, API около 733/800%, LB далёк от квоты. +Это сочетание CPU приложения и ожидания клиента под блокировкой чата; рост очереди +LB — следствие, а не доказательство насыщения CPU балансировщика. Контроль пулов и +кэширование CTE проверяют эту интерпретацию. Все 90 точек приняты без помех. diff --git a/openspec/changes/remeasure-lab3-scaling/tasks.md b/openspec/changes/remeasure-lab3-scaling/tasks.md index 2716c00b..515b264a 100644 --- a/openspec/changes/remeasure-lab3-scaling/tasks.md +++ b/openspec/changes/remeasure-lab3-scaling/tasks.md @@ -5,12 +5,12 @@ ## 2. Базовые серии SQLAlchemy -- [ ] 2.1 Снять хост и серию A: 1×4/2×2/4×1, 3 повтора каждой точки. -- [ ] 2.2 Снять серию B: 1×1/2×1/8×1, найти и объяснить насыщенное звено. +- [x] 2.1 Снять хост и серию A: 1×4/2×2/4×1, 3 повтора каждой точки. +- [x] 2.2 Снять серию B: 1×1/2×1/8×1, найти и объяснить насыщенное звено. ## 3. Улучшение messaging -- [ ] 3.1 Кэшировать неизменный CTE; проверить API, гонки, TTL, rollback и изоляцию параметров. +- [x] 3.1 Кэшировать неизменный CTE; проверить API, гонки, TTL, rollback и изоляцию параметров. - [ ] 3.2 Провести ≥3 чередующихся повтора before/after на одной конфигурации. - [ ] 3.3 Провести контроль пулов/maxconn и числа процессов на одинаковых CPU. diff --git a/scripts/scaling.py b/scripts/scaling.py index 793c8c7c..7fcfd929 100755 --- a/scripts/scaling.py +++ b/scripts/scaling.py @@ -259,6 +259,8 @@ def provenance(folder, args, env): "project": args.project, "command": sys.argv, "api_image": env.get("API_IMAGE", "compose default"), + "warmup_seconds": args.warmup, + "duration_seconds": args.duration, "containers": args.deployment, } (folder / "metadata.json").write_text(json.dumps(metadata, ensure_ascii=False) + "\n") @@ -344,6 +346,7 @@ def step(base, rate, repeat, args, env): or row["gap_s"] > args.max_gap or wall > args.warmup + args.duration + args.max_overhead or row["cpu_k6"] >= 95 * float(env.get("K6_CPUS", "2")) + or any(row[f"cpu_{name}"] <= 0 for name in ("vm", "api", "postgres", "k6")) or not rows(folder / "pg_activity.csv", start, end) or not rows(folder / "lb_stats.csv", start, end) ) diff --git a/src/messenger/messaging/infrastructure/postgres.py b/src/messenger/messaging/infrastructure/postgres.py index 99445f42..c40a84b6 100644 --- a/src/messenger/messaging/infrastructure/postgres.py +++ b/src/messenger/messaging/infrastructure/postgres.py @@ -18,6 +18,7 @@ from sqlalchemy import ( bindparam, func, insert, + lambda_stmt, literal, select, tuple_, @@ -173,7 +174,7 @@ def post_message() -> Select: ).add_cte(touched) -POST_MESSAGE = post_message() +POST_MESSAGE = lambda_stmt(post_message, enable_tracking=False) def member_ids() -> Label[Sequence[UUID]]: diff --git a/student-tests/api/test_idempotency.py b/student-tests/api/test_idempotency.py index 4c883b10..8ea53d95 100644 --- a/student-tests/api/test_idempotency.py +++ b/student-tests/api/test_idempotency.py @@ -95,12 +95,22 @@ async def test_key_is_scoped_by_user_and_chat(client: AsyncClient) -> None: bob = await register(client, "bob") chat = await make_chat(client, alice, bob) other = await make_chat(client, alice) - ids = { - (await send(client, chat, alice, "алиса", "shared-1")).json()["id"], - (await send(client, chat, bob, "боб", "shared-1")).json()["id"], - (await send(client, other, alice, "другой чат", "shared-1")).json()["id"], - } + cases = [(chat, alice, "алиса"), (chat, bob, "боб"), (other, alice, "другой чат")] + messages = [] + for target, user, text in cases: + response = await send(client, target, user, text, "shared-1") + assert response.status_code == 201 + message = response.json() + assert (message["chat_id"], message["sender_id"], message["text"]) == ( + target["id"], + user["id"], + text, + ) + messages.append(message) + ids = {message["id"] for message in messages} assert len(ids) == 3 + for (target, user, text), message in zip(cases, messages, strict=True): + assert (await send(client, target, user, text, "shared-1")).json() == message async def test_replay_checks_session_and_access_again(client: AsyncClient) -> None: diff --git a/tools/plot_track1.py b/tools/plot_track1.py index b208a19d..a04d08b8 100644 --- a/tools/plot_track1.py +++ b/tools/plot_track1.py @@ -2,6 +2,8 @@ import argparse import collections import csv import statistics +import subprocess +import time from pathlib import Path import matplotlib.pyplot as plt @@ -13,8 +15,13 @@ USL_POINTS = 3 MIN_REPEATS = 3 COMPONENTS = (("api", "API"), ("postgres", "PostgreSQL"), ("lb", "HAProxy"), ("k6", "k6")) INK, MUTED, GRID, SURFACE = "#0b0b0b", "#52514e", "#e4e3df", "#fcfcfb" +STAMP = ( + f"date: {time.strftime('%FT%TZ', time.gmtime())} commit: " + + subprocess.check_output(["/usr/bin/env", "git", "rev-parse", "HEAD"], text=True).strip() +) SUMMARY = ( - "series variant config rates capacity_rate capacity_ok_rps capacity_p95_ms capacity_p99_ms " + "series variant config rates capacity_rate capacity_label " + "capacity_ok_rps capacity_p95_ms capacity_p99_ms " "peak_rate peak_ok_rps peak_ok_min peak_ok_max peak_p95_ms peak_failed_ratio peak_dropped " "cpu_api cpu_api_limit cpu_postgres cpu_lb cpu_k6 pg_active_avg pg_lock_avg pg_idle_tx_avg " "pg_idle_avg pool_total pg_backends_max lb_qcur_max" @@ -114,6 +121,9 @@ def summarize(key, rates): for field, name in (("ok_rps", "ok_rps"), ("p95_ms", "p95_ms"), ("p99_ms", "p99_ms")): row[f"capacity_{name}"] = med(rates[capacity], field)[0] row["capacity_rate"] = capacity + row["capacity_label"] = ( + f"≥{capacity}" if capacity in (min(rates), max(rates)) else str(capacity) + ) return row @@ -123,13 +133,21 @@ def direct_label(axis, x, y, text): ) -def lines(axis, keys, groups, field): - for key in keys: +def save(figure, path): + figure.tight_layout() + figure.savefig(path, dpi=150, metadata={"Description": STAMP}) + plt.close(figure) + + +def lines(axis, keys, groups, field, categorical=False): + for index, key in enumerate(keys): rates = sorted(groups[key]) values = [med(groups[key][rate], field) for rate in rates] look = VARIANTS.get(key[1]) or style(key[2]) + if categorical: + look = {"color": PALETTE[index], "marker": MARKERS[index]} axis.errorbar( - rates, + [index] if categorical else rates, [v[0] for v in values], yerr=[[v[1] for v in values], [v[2] for v in values]], capsize=3, @@ -138,28 +156,34 @@ def lines(axis, keys, groups, field): markeredgewidth=1.5, **look, ) - direct_label(axis, rates[-1], values[-1][0], label(key)) + if not categorical: + direct_label(axis, rates[-1], values[-1][0], label(key)) + if categorical: + axis.set_xticks(range(len(keys)), [label(k) for k in keys], rotation=25, ha="right") def throughput(series, keys, groups, out): figure, axes = plt.subplots(1, 3, figsize=(17, 4.8)) top = max(rate for key in keys for rate in groups[key]) - axes[0].plot( - [0, top], [0, top], color=MUTED, linestyle="--", linewidth=1, label="RPS = нагрузка" - ) - lines(axes[0], keys, groups, "ok_rps") + categorical = len({rate for key in keys for rate in groups[key]}) == 1 + if not categorical: + axes[0].plot( + [0, top], [0, top], color=MUTED, linestyle="--", linewidth=1, label="RPS = нагрузка" + ) + lines(axes[0], keys, groups, "ok_rps", categorical) axes[0].set_title("Успешные RPS") for axis, field, title in ((axes[1], "p95_ms", "p95, мс"), (axes[2], "p99_ms", "p99, мс")): - lines(axis, keys, groups, field) + lines(axis, keys, groups, field, categorical) axis.set_yscale("log") axis.set_title(title) for axis in axes: - axis.set_xlabel("Предложенная нагрузка, запросов/с") - axis.legend(loc="upper left", fontsize=9) - figure.suptitle(f"{series}: медиана 3 повторов, усы — min–max", color=INK) - figure.tight_layout() - figure.savefig(out / f"{series}-throughput.png", dpi=150) - plt.close(figure) + axis.set_xlabel( + f"Вариант при {top} запросах/с" if categorical else "Предложенная нагрузка, запросов/с" + ) + if not categorical: + axis.legend(loc="upper left", fontsize=9) + figure.suptitle(f"{series}: медиана повторов, усы — min–max", color=INK) + save(figure, out / f"{series}-throughput.png") def cpu(series, keys, groups, out): @@ -176,9 +200,7 @@ def cpu(series, keys, groups, out): axis.set_title(f"CPU, % ядра — {label(key)}") axis.set_xlabel("Предложенная нагрузка, запросов/с") axis.legend(fontsize=8, loc="upper left") - figure.tight_layout() - figure.savefig(out / f"{series}-cpu.png", dpi=150) - plt.close(figure) + save(figure, out / f"{series}-cpu.png") def usl(points): @@ -248,9 +270,7 @@ def scaling(rows, out): axis.set_title("Пропускная способность от ресурсов API") axis.set_xticks(cpus) axis.legend(loc="upper left", fontsize=9) - figure.tight_layout() - figure.savefig(out / "scaling.png", dpi=150) - plt.close(figure) + save(figure, out / "scaling.png") return fit @@ -288,13 +308,12 @@ def utilisation(rows, out): axis.set_ylim(0, 115) axis.set_title("Загрузка звеньев в точке пика успешных RPS", pad=34) axis.legend(fontsize=9, ncols=4, loc="lower left", bbox_to_anchor=(0, 1.0)) - figure.tight_layout() - figure.savefig(out / "utilisation.png", dpi=150) - plt.close(figure) + save(figure, out / "utilisation.png") def write(rows, out, fit): with (out / "capacity.csv").open("w", newline="") as handle: + handle.write(f"# {STAMP}\n") writer = csv.DictWriter(handle, fieldnames=SUMMARY, extrasaction="ignore") writer.writeheader() writer.writerows(rows) @@ -305,7 +324,7 @@ def write(rows, out, fit): lines_ = [head, "|" + " --- |" * 11] for r in rows: lines_.append( - f"| {r['series']} | {r['variant']} | {r['config']} | {r.get('capacity_rate', '—')} | " + f"| {r['series']} | {r['variant']} | {r['config']} | {r.get('capacity_label', '—')} | " f"{r.get('capacity_p95_ms', '—')} | {r['peak_ok_rps']:.0f} " f"[{r['peak_ok_min']:.0f}–{r['peak_ok_max']:.0f}] | {r['peak_rate']} | " f"{r['cpu_api']:.0f} / {r['cpu_api_limit']} | {r['cpu_postgres']:.0f} | " @@ -315,9 +334,10 @@ def write(rows, out, fit): _, sigma, kappa, peak = fit best = f"{peak:.1f}" if peak else "—" lines_.append( - f"\nUSL по серии B: σ = {sigma:.4f}, κ = {kappa:.5f}, пик при N* = {best} CPU" + f"\nОписательная USL-аппроксимация серии B: σ = {sigma:.4f}, κ = {kappa:.5f}, " + f"модельный N* = {best} CPU. Это не измеренный оптимум и не доказательство причины." ) - (out / "capacity.md").write_text("\n".join(lines_) + "\n") + (out / "capacity.md").write_text(STAMP + "\n\n" + "\n".join(lines_) + "\n") print("\n".join(lines_))