forked from mtusi2026/mtusi2026lab2
142 lines
8.1 KiB
Python
142 lines
8.1 KiB
Python
"""Трек 2: два индекса и OFFSET/keyset на 100 000 сообщений.
|
|
|
|
Только синтетические данные в отдельной схеме. Не меняет таблицы API.
|
|
EXPLAIN включает BUFFERS, планировщику не навязывается тип плана.
|
|
"""
|
|
import json
|
|
import os
|
|
import platform
|
|
import random
|
|
import statistics
|
|
import time
|
|
from datetime import datetime, timedelta, timezone
|
|
from pathlib import Path
|
|
from uuid import uuid4
|
|
|
|
import psycopg
|
|
from psycopg import sql
|
|
from psycopg.rows import dict_row
|
|
from src.config import DATABASE_URL
|
|
|
|
SEED = 42
|
|
SIZES = [70000, 20000, 9000, 900, 100]
|
|
REPEATS = 30
|
|
|
|
|
|
def percentile(values, p):
|
|
return sorted(values)[max(0, int(len(values) * p + 0.9999) - 1)]
|
|
|
|
|
|
def measure(db, query, params):
|
|
for _ in range(5):
|
|
db.execute(query, params).fetchall()
|
|
raw = []
|
|
for _ in range(REPEATS):
|
|
start = time.perf_counter()
|
|
db.execute(query, params).fetchall()
|
|
raw.append((time.perf_counter() - start) * 1000)
|
|
plan = db.execute('EXPLAIN (ANALYZE, BUFFERS, FORMAT JSON) ' + query, params).fetchone()[0][0]
|
|
return {'latency_ms': raw, 'p50_ms': statistics.median(raw),
|
|
'p95_ms': percentile(raw, .95), 'p99_ms': percentile(raw, .99), 'plan': plan}
|
|
|
|
|
|
def main():
|
|
schema = 'research_' + uuid4().hex[:12]
|
|
rng = random.Random(SEED)
|
|
out = {'seed': SEED, 'schema': schema, 'message_count': sum(SIZES), 'sizes': SIZES,
|
|
'repeats': REPEATS, 'warmup': 5, 'concurrency': 1,
|
|
'platform': platform.platform(), 'cpu_count': os.cpu_count(), 'results': []}
|
|
Path('evidence').mkdir(exist_ok=True)
|
|
with psycopg.connect(DATABASE_URL, autocommit=True) as db:
|
|
out['postgres_version'] = db.execute('SHOW server_version').fetchone()[0]
|
|
out['settings'] = {k: db.execute('SHOW ' + k).fetchone()[0] for k in
|
|
['shared_buffers', 'work_mem', 'fsync', 'synchronous_commit']}
|
|
db.execute(sql.SQL('CREATE SCHEMA {}').format(sql.Identifier(schema)))
|
|
db.execute(sql.SQL('SET search_path TO {}').format(sql.Identifier(schema)))
|
|
# Точная схема приложения; другой search_path изолирует эксперимент.
|
|
db.execute(Path('migrations/001_initial.sql').read_text())
|
|
user = uuid4()
|
|
db.execute('INSERT INTO users(id,username,display_name,password_hash) VALUES (%s,%s,%s,%s)',
|
|
(user, 'research_user', 'Synthetic research', 'not-a-login-account'))
|
|
chats = [uuid4() for _ in SIZES]
|
|
start_date = datetime(2026, 1, 1, tzinfo=timezone.utc)
|
|
for chat, size in zip(chats, SIZES):
|
|
db.execute('INSERT INTO chats(id,title) VALUES (%s,%s)', (chat, 'Synthetic ' + str(size)))
|
|
db.execute('INSERT INTO chat_members VALUES (%s,%s)', (chat, user))
|
|
with db.cursor().copy('COPY messages(id,chat_id,sender_id,text,created_at) FROM STDIN') as copy:
|
|
for n in range(size):
|
|
copy.write_row((uuid4(), chat, user, 'message ' + str(n) + ' x' * (n % 80),
|
|
start_date + timedelta(microseconds=n)))
|
|
samples = []
|
|
for chat in chats:
|
|
for fraction in [0, .5, .9]:
|
|
count = db.execute('SELECT count(*) FROM messages WHERE chat_id=%s', (chat,)).fetchone()[0]
|
|
offset = int((count - 1) * fraction)
|
|
cursor = db.execute('SELECT created_at,id FROM messages WHERE chat_id=%s '
|
|
'ORDER BY created_at,id OFFSET %s LIMIT 1', (chat, offset)).fetchone()
|
|
samples.append((chat, offset, cursor))
|
|
db.execute('DROP INDEX messages_history_idx')
|
|
# Порядок фиксирован, но обе конфигурации прогреваются и имеют 30 повторов.
|
|
for name, columns in [('chat_only', 'chat_id'), ('history', 'chat_id,created_at,id')]:
|
|
db.execute('CREATE INDEX research_candidate ON messages(' + columns + ')')
|
|
db.execute('ANALYZE messages')
|
|
index_bytes = db.execute("SELECT pg_relation_size('research_candidate')").fetchone()[0]
|
|
queries = list(samples)
|
|
rng.shuffle(queries)
|
|
for chat, offset, cursor in queries:
|
|
for method in ['offset', 'keyset']:
|
|
query = 'SELECT * FROM messages WHERE chat_id=%s'
|
|
if method == 'keyset':
|
|
query += ' AND (created_at,id)>(%s,%s) ORDER BY created_at,id LIMIT 51'
|
|
params = (chat, *cursor)
|
|
else:
|
|
# Cursor указывает последнюю уже прочитанную строку.
|
|
query += ' ORDER BY created_at,id OFFSET %s LIMIT 51'
|
|
params = (chat, offset + 1)
|
|
result = measure(db, query, params)
|
|
result.update(index=name, method=method, chat_size=SIZES[chats.index(chat)],
|
|
offset=offset, index_bytes=index_bytes)
|
|
out['results'].append(result)
|
|
write_ms = []
|
|
for batch in range(5):
|
|
# Цена 1000 записей + COMMIT; затем откат данных отдельной очисткой.
|
|
# Исходная история остаётся 100 000 для следующего индекса.
|
|
batch_ids = [uuid4() for _ in range(1000)]
|
|
begin = time.perf_counter()
|
|
with db.transaction():
|
|
with db.cursor().copy('COPY messages(id,chat_id,sender_id,text,created_at) FROM STDIN') as copy:
|
|
for n, mid in enumerate(batch_ids):
|
|
copy.write_row((mid, chats[n % len(chats)], user, 'write-cost ' + 'x' * 128,
|
|
start_date + timedelta(days=1, microseconds=n)))
|
|
write_ms.append((time.perf_counter() - begin) * 1000)
|
|
db.execute('DELETE FROM messages WHERE id=ANY(%s)', (batch_ids,))
|
|
out.setdefault('writes', {})[name] = {'batch_rows': 1000, 'batch_ms': write_ms,
|
|
'median_ms': statistics.median(write_ms),
|
|
'index_bytes': index_bytes}
|
|
db.execute('DROP INDEX research_candidate')
|
|
db.execute('VACUUM ANALYZE messages')
|
|
db.execute('CREATE INDEX messages_history_idx ON messages(chat_id,created_at,id)')
|
|
with db.cursor(row_factory=dict_row) as cur:
|
|
out['counts'] = {}
|
|
for table in ['users', 'chats', 'chat_members', 'messages', 'sessions']:
|
|
out['counts'][table] = db.execute('SELECT count(*) FROM ' + table).fetchone()[0]
|
|
cur.execute('SELECT * FROM messages ORDER BY created_at,id LIMIT 5')
|
|
out['sample_messages'] = [dict(row) for row in cur.fetchall()]
|
|
Path('evidence/track2_raw.json').write_text(json.dumps(out, ensure_ascii=False, indent=2, default=str))
|
|
lines = ['# Трек 2: измерения', '', 'Исходных сообщений: 100000. Прогрев: 5; повторов: 30.', '',
|
|
'| Чат | Позиция | Индекс | Запрос | p50, мс | p95, мс |',
|
|
'|---:|---:|---|---|---:|---:|']
|
|
for row in sorted(out['results'], key=lambda r: (r['chat_size'], r['offset'], r['index'], r['method'])):
|
|
lines.append(f"| {row['chat_size']} | {row['offset']} | {row['index']} | {row['method']} | "
|
|
f"{row['p50_ms']:.3f} | {row['p95_ms']:.3f} |")
|
|
lines.extend(['', 'Цена записи: 5 транзакций по 1000 строк (COPY + COMMIT).', '',
|
|
'| Индекс | Медиана партии, мс | Размер, байт |', '|---|---:|---:|'])
|
|
for name, row in out['writes'].items():
|
|
lines.append(f"| {name} | {row['median_ms']:.3f} | {row['index_bytes']} |")
|
|
Path('evidence/track2_summary.md').write_text('\n'.join(lines) + '\n')
|
|
print('PASS: measured two indexes, two queries, write cost; raw evidence saved')
|
|
|
|
|
|
if __name__ == '__main__':
|
|
main()
|