Оптимизация производительности вставки и чтения для S3

① Получить и разобрать следующую порцию данных и сформировать из неё блок данных в памяти (по одному на ключ партиционирования).

② Записать блок в новую часть на хранилище.

Перейти к ① 

① Decide which parts to merge next, and load these parts as blocks into memory.

② Merge the loaded blocks in memory into a larger block.

③ Write the merged block into a new part on disk.

Go to ①

① Get the next portion of unprocessed file data (portion size is based on the configured block size) and create an in-memory data block from it.

② Write the block into a new part on storage.

Go to ①. 

-- Топ пользователей по именам
SELECT
    OwnerDisplayName,
    count() AS num_posts
FROM s3('https://datasets-documentation.s3.eu-west-3.amazonaws.com/stackoverflow/parquet/posts/by_month/*.parquet')
WHERE OwnerDisplayName NOT IN ('', 'anon')
GROUP BY OwnerDisplayName
ORDER BY num_posts DESC
LIMIT 5

┌─OwnerDisplayName─┬─num_posts─┐
│ user330315       │     10344 │
│ user4039065      │      5316 │
│ user149341       │      4102 │
│ user529758       │      3700 │
│ user3559349      │      3068 │
└──────────────────┴───────────┘

5 rows in set. Elapsed: 3.013 sec. Processed 59.82 million rows, 24.03 GB (19.86 million rows/s., 7.98 GB/s.)
Peak memory usage: 603.64 MiB.

-- Загрузка в таблицу posts
INSERT INTO posts SELECT *
FROM s3('https://datasets-documentation.s3.eu-west-3.amazonaws.com/stackoverflow/parquet/posts/by_month/*.parquet')

0 rows in set. Elapsed: 191.692 sec. Processed 59.82 million rows, 24.03 GB (312.06 thousand rows/s., 125.37 MB/s.)

SELECT
    OwnerDisplayName,
    count() AS num_posts
FROM s3('https://datasets-documentation.s3.eu-west-3.amazonaws.com/stackoverflow/parquet/posts/by_month/*.parquet')
WHERE OwnerDisplayName NOT IN ('', 'anon')
GROUP BY OwnerDisplayName
ORDER BY num_posts DESC
LIMIT 5
SETTINGS max_threads = 16

┌─OwnerDisplayName─┬─num_posts─┐
│ user330315       │     10344 │
│ user4039065      │      5316 │
│ user149341       │      4102 │
│ user529758       │      3700 │
│ user3559349      │      3068 │
└──────────────────┴───────────┘

5 rows in set. Elapsed: 1.505 sec. Processed 59.82 million rows, 24.03 GB (39.76 million rows/s., 15.97 GB/s.)
Peak memory usage: 178.58 MiB.

SETTINGS max_threads = 32

5 rows in set. Elapsed: 0.779 sec. Processed 59.82 million rows, 24.03 GB (76.81 million rows/s., 30.86 GB/s.)
Peak memory usage: 369.20 MiB.

SETTINGS max_threads = 64

5 rows in set. Elapsed: 0.674 sec. Processed 59.82 million rows, 24.03 GB (88.81 million rows/s., 35.68 GB/s.)
Peak memory usage: 639.99 MiB.

• max_insert_threads: выберите ~ половину доступных ядер CPU для потоков вставки (чтобы оставить достаточно ядер для фоновых слияний)

• peak_memory_usage_in_bytes: выберите планируемый пиковый объём потребления памяти: либо весь доступный RAM (если загрузка данных выполняется изолированно), либо половину или меньше (чтобы оставить ресурсы для других параллельных задач)

Затем:
min_insert_block_size_bytes = peak_memory_usage_in_bytes / (~3 * max_insert_threads)

INSERT INTO posts SELECT *
FROM s3('https://datasets-documentation.s3.eu-west-3.amazonaws.com/stackoverflow/parquet/posts/by_month/*.parquet') SETTINGS min_insert_block_size_rows=0, max_insert_threads=4, min_insert_block_size_bytes=2863311530

0 rows in set. Elapsed: 128.566 sec. Processed 59.82 million rows, 24.03 GB (465.28 thousand rows/s., 186.92 MB/s.)

INSERT INTO posts SELECT *
FROM s3('https://datasets-documentation.s3.eu-west-3.amazonaws.com/stackoverflow/parquet/posts/by_month/*.parquet') SETTINGS min_insert_block_size_rows=0, max_insert_threads=8, min_insert_block_size_bytes=2863311530

0 rows in set. Elapsed: 67.294 sec. Processed 59.82 million rows, 24.03 GB (888.93 thousand rows/s., 357.12 MB/s.)

SELECT
    OwnerDisplayName,
    count() AS num_posts
FROM s3('https://datasets-documentation.s3.eu-west-3.amazonaws.com/stackoverflow/parquet/posts/by_month/*.parquet')
WHERE OwnerDisplayName NOT IN ('', 'anon')
GROUP BY OwnerDisplayName
ORDER BY num_posts DESC
LIMIT 5
SETTINGS max_threads = 92

5 rows in set. Elapsed: 0.421 sec. Processed 59.82 million rows, 24.03 GB (142.08 million rows/s., 57.08 GB/s.)

SELECT
    OwnerDisplayName,
    count() AS num_posts
FROM s3Cluster('default', 'https://datasets-documentation.s3.eu-west-3.amazonaws.com/stackoverflow/parquet/posts/by_month/*.parquet')
WHERE OwnerDisplayName NOT IN ('', 'anon')
GROUP BY OwnerDisplayName
ORDER BY num_posts DESC
LIMIT 5
SETTINGS max_threads = 16

┌─OwnerDisplayName─┬─num_posts─┐
│ user330315       │     10344 │
│ user4039065      │      5316 │
│ user149341       │      4102 │
│ user529758       │      3700 │
│ user3559349      │      3068 │
└──────────────────┴───────────┘

5 rows in set. Elapsed: 0.622 sec. Processed 59.82 million rows, 24.03 GB (96.13 million rows/s., 38.62 GB/s.)
Peak memory usage: 176.74 MiB.

INSERT INTO posts SELECT *
FROM s3Cluster('default', 'https://datasets-documentation.s3.eu-west-3.amazonaws.com/stackoverflow/parquet/posts/by_month/*.parquet') SETTINGS min_insert_block_size_rows=0, max_insert_threads=4, min_insert_block_size_bytes=2863311530

0 rows in set. Elapsed: 171.202 sec. Processed 59.82 million rows, 24.03 GB (349.41 thousand rows/s., 140.37 MB/s.)

INSERT INTO posts
SELECT *
FROM s3Cluster('default', 'https://datasets-documentation.s3.eu-west-3.amazonaws.com/stackoverflow/parquet/posts/by_month/*.parquet')
SETTINGS parallel_distributed_insert_select = 2, min_insert_block_size_rows=0, max_insert_threads=4, min_insert_block_size_bytes=2863311530

0 rows in set. Elapsed: 54.571 sec. Processed 59.82 million rows, 24.03 GB (1.10 million rows/s., 440.38 MB/s.)
Peak memory usage: 11.75 GiB.

INSERT INTO posts
SETTINGS parallel_distributed_insert_select = 2, min_insert_block_size_rows = 0, max_insert_threads = 4, min_insert_block_size_bytes = 2863311530, insert_deduplicate = 0
SELECT *
FROM s3Cluster('default', 'https://datasets-documentation.s3.eu-west-3.amazonaws.com/stackoverflow/parquet/posts/by_month/*.parquet')
SETTINGS parallel_distributed_insert_select = 2, min_insert_block_size_rows = 0, max_insert_threads = 4, min_insert_block_size_bytes = 2863311530, insert_deduplicate = 0

0 rows in set. Elapsed: 52.992 sec. Processed 59.82 million rows, 24.03 GB (1.13 million rows/s., 453.50 MB/s.)
Peak memory usage: 26.57 GiB.

SELECT *
FROM s3Cluster('default', 'https://datasets-documentation.s3.eu-west-3.amazonaws.com/stackoverflow/parquet/posts/by_month/*.parquet')
SETTINGS parallel_distributed_insert_select = 2, min_insert_block_size_rows = 0, max_insert_threads = 4, min_insert_block_size_bytes = 2863311530, insert_deduplicate = 0, optimize_on_insert = 0

0 rows in set. Elapsed: 49.688 sec. Processed 59.82 million rows, 24.03 GB (1.20 million rows/s., 483.66 MB/s.)