aka "Data skipping indices"
Collect a summary of column/expression values for every N granules.
Use this summaries to skip data while reading.
Indices are available for MergeTree family of table engines.
SET allow_experimental_data_skipping_indices = 1;
CREATE TABLE table
(...
INDEX name expr TYPE type(params...) GRANULARITY n
...)
ALTER TABLE ... ADD INDEX name expr
TYPE type(params...) GRANULARITY n
ALTER TABLE ... DROP INDEX name
minmax
— summary is just min/max boundaries of values;
— use when values are correlated to table order;
or distributed locally; or sparse;
set(k)
— summary is a set of all distinct values, but not larger than k;
— use when values are sparse or have low cardinality;
— reasonable values of k is about hundred;
Used for comparison and IN operators.
Full text search indices (highly experimental)
ngrambf_v1(chars, size, hashes, seed)
tokenbf_v1(size, hashes, seed)
Used for equals comparison, IN and LIKE.
SELECT count()
FROM test.hits
WHERE URLDomain LIKE 'aena.es'
┌─count()─┐
│ 1 │
└─────────┘
Processed 8.87 million rows
SET allow_experimental_data_skipping_indices = 1;
ALTER TABLE test.hits
ADD INDEX
domain_index URLDomain
TYPE set(1000) GRANULARITY 1;
OPTIMIZE TABLE test.hits FINAL;
SELECT count()
FROM test.hits
WHERE URLDomain LIKE 'aena.es'
┌─count()─┐
│ 1 │
└─────────┘
Processed 65.54 thousand rows
col type CODEC(codecs...)
Available codecs:
— LZ4 (default);
— ZSTD; — level can be specified: ZSTD(1);
— LZ4HC; — level can be specified;
— NONE;
— Delta(N); — N is the size of the data type in bytes.
Codecs can be chained together:
time DateTime CODEC(Delta(4), LZ4)
— Delta is not a compression itself, it must be chained with a second codec.
Per-column codecs have priority over <compression> config settings.
SELECT
name,
type,
formatReadableSize(data_compressed_bytes) AS compressed,
formatReadableSize(data_uncompressed_bytes) AS uncompressed,
data_uncompressed_bytes / data_compressed_bytes AS ratio,
compression_codec
FROM system.columns
WHERE (database = 'test') AND (table = 'hits')
ORDER BY data_compressed_bytes DESC
LIMIT 10
┌─name────────────┬─type─────┬─compressed─┬─uncompressed─┬──────────────ratio─┬─compression_codec─┐
│ Referer │ String │ 180.19 MiB │ 582.99 MiB │ 3.2353881463220975 │ │
│ URL │ String │ 128.93 MiB │ 660.58 MiB │ 5.123600238954646 │ │
│ Title │ String │ 95.29 MiB │ 595.01 MiB │ 6.244488505685867 │ │
│ WatchID │ UInt64 │ 67.28 MiB │ 67.70 MiB │ 1.0062751884416956 │ │
│ URLHash │ UInt64 │ 37.09 MiB │ 67.70 MiB │ 1.8254645825020759 │ │
│ ClientEventTime │ DateTime │ 31.42 MiB │ 33.85 MiB │ 1.0772947535816229 │ │
│ EventTime │ DateTime │ 31.40 MiB │ 33.85 MiB │ 1.0780959105750834 │ │
│ UTCEventTime │ DateTime │ 31.39 MiB │ 33.85 MiB │ 1.0783175064258996 │ │
│ HID │ UInt32 │ 28.28 MiB │ 33.85 MiB │ 1.19709852035762 │ │
│ RefererHash │ UInt64 │ 27.68 MiB │ 67.70 MiB │ 2.445798559204409 │ │
└─────────────────┴──────────┴────────────┴──────────────┴────────────────────┴───────────────────┘
ALTER TABLE test.hits
MODIFY COLUMN ClientEventTime CODEC(Delta(4), LZ4)
Changes are applied lazily: only for new data and while merging.
ALTER TABLE test.hits
UPDATE ClientEventTime = ClientEventTime WHERE 1
— a trick to rewrite column data on disk.
— also executed in background, look at system.mutations table.
SELECT
name,
type,
formatReadableSize(data_compressed_bytes) AS compressed,
formatReadableSize(data_uncompressed_bytes) AS uncompressed,
data_uncompressed_bytes / data_compressed_bytes AS ratio,
compression_codec
FROM system.columns
WHERE (database = 'test') AND (table = 'hits') AND (name = 'ClientEventTime')
ORDER BY data_compressed_bytes DESC
LIMIT 10
┌─name────────────┬─type─────┬─compressed─┬─uncompressed─┬──────────────ratio─┬─compression_codec────┐
│ ClientEventTime │ DateTime │ 19.47 MiB │ 33.85 MiB │ 1.7389218149308554 │ CODEC(Delta(4), LZ4) │
└─────────────────┴──────────┴────────────┴──────────────┴────────────────────┴──────────────────────┘
ALTER TABLE test.hits
MODIFY COLUMN
ClientEventTime CODEC(Delta(4), ZSTD),
UPDATE ClientEventTime = ClientEventTime WHERE 1
┌─name────────────┬─type─────┬─compressed─┬─uncompressed─┬─────────────ratio─┬─compression_codec────────┐
│ ClientEventTime │ DateTime │ 14.00 MiB │ 33.85 MiB │ 2.417489322394391 │ CODEC(Delta(4), ZSTD(1)) │
└─────────────────┴──────────┴────────────┴──────────────┴───────────────────┴──────────────────────────┘
Just replace String to LowCardinality(String)
for string fields with low number of unique values.
... and it will magically work faster.
For high cardinality fields it will work fine but pointless.
Examples:
— city name — ok;
— domain of URL — ok;
— search phrase — bad;
— URL — bad;
SELECT count()
FROM hits_333
WHERE URLDomain LIKE 'aena.es'
┌─count()─┐
│ 101 │
└─────────┘
1 rows in set. Elapsed: 0.446 sec.
Processed 333.36 million rows, 7.32 GB
(747.87 million rows/s., 16.43 GB/s.)
ALTER TABLE hits_333
MODIFY COLUMN
URLDomain LowCardinality(String)
Ok.
0 rows in set. Elapsed: 16.228 sec.
SELECT count()
FROM hits_333
WHERE URLDomain LIKE 'aena.es'
┌─count()─┐
│ 101 │
└─────────┘
1 rows in set. Elapsed: 0.244 sec.
Processed 333.36 million rows, 1.72 GB
(1.37 billion rows/s., 7.04 GB/s.)
Two times faster!
Multiple substring search
Multiple regexp search
Fuzzy string comparison and search
Fuzzy regexp match
SELECT count()
FROM hits_100m
WHERE multiSearchAny(URL,
['chelyabinsk.74.ru', 'doctor.74.ru', 'transport.74.ru',
'm.74.ru', 'chel.74.ru', 'afisha.74.ru', 'diplom.74.ru',
'//chel.ru', 'chelyabinsk.ru', 'cheldoctor.ru'])
— multiSearchAny
— multiSearchFirstPosition
— multiSearchFirstIndex
— multiSearchAllPositions
+ -UTF8, -CaseInsensitive, -CaseInsensitiveUTF8
— multiMatchAny
— multiMatchAnyIndex
— multiFuzzyMatchAny
— multiFuzzyMatchAnyIndex
— ngramDistance
+ -UTF8, -CaseInsensitive, -CaseInsensitiveUTF8
SELECT DISTINCT
SearchPhrase,
ngramDistance(SearchPhrase, 'clickhouse') AS dist
FROM hits_100m_single
ORDER BY dist ASC
LIMIT 10
┌─SearchPhrase────┬───────dist─┐
│ tickhouse │ 0.23076923 │
│ clockhouse │ 0.42857143 │
│ house │ 0.5555556 │
│ clickhomecyprus │ 0.57894737 │
│ 1click │ 0.6 │
│ uhouse │ 0.6 │
│ teakhouse.ru │ 0.625 │
│ teakhouse.com │ 0.64705884 │
│ madhouse │ 0.6666667 │
│ funhouse │ 0.6666667 │
└─────────────────┴────────────┘
10 rows in set. Elapsed: 1.267 sec. Processed 100.00 million rows, 1.52 GB
(78.92 million rows/s., 1.20 GB/s.)
(contributed by TouTiao/ByteDance)
SELECT * FROM hdfs(
'hdfs://hdfs1:9000/file',
'TSV', 'id UInt64, text String);
INSERT INTO TABLE FUNCTION hdfs(
'hdfs://hdfs1:9000/file',
'TSV', 'id UInt64, text String) VALUES ...
CREATE TABLE (...) ENGINE =
HDFS('hdfs://hdfs1:9000/file', 'TSV');
Drawbacks: not all authentication methods supported.
Protobuf
— efficient implementation, no excessive copies/allocations
(ClickHouse style);
— transparent type conversions between Proto's and ClickHouse types (UInt8, Int64, DateTime <-> sint64, uint64, sint32, uint32, String <-> bytes, string, etc.);
— support for Nested types via repeated Messages or parallel repeated fields;
— format_schema setting must be specified.
Parquet
— columnar format; naturally implemented without unpacking of columns;
— transparent type conversions also supported.
IPv4 and IPv6 data types
(contributed by Altinity)
SELECT DISTINCT ClientIP6
FROM test.hits
LIMIT 10
┌─ClientIP6─┐
│
│ ��m��� │
│ �� � │
│ ���9EY │
│ ���9�� │
│ ��]�� │
│ ��T/�� │
│ ��m�(� │
│ ��_N( │
└───────────┘
SELECT DISTINCT IPv6NumToString(ClientIP6)
FROM test.hits
LIMIT 10
┌─IPv6NumToString(ClientIP6)─┐
│ ::ffff:128.70.160.13 │
│ ::ffff:109.167.145.182 │
│ ::ffff:46.8.32.249 │
│ ::ffff:178.57.69.89 │
│ ::ffff:178.57.159.237 │
│ ::ffff:2.93.132.186 │
│ ::ffff:84.47.183.210 │
│ ::ffff:109.184.40.252 │
│ ::ffff:95.78.40.21 │
└────────────────────────────┘
ALTER TABLE test.hits
MODIFY COLUMN
ClientIPv6 IPv6
SELECT DISTINCT ClientIP6
FROM test.hits
LIMIT 10
┌─ClientIP6──────────────┐
│ ::ffff:93.77.44.100 │
│ ::ffff:176.59.183.15 │
│ ::ffff:46.0.180.32 │
│ ::ffff:46.0.180.50 │
│ ::ffff:178.69.88.247 │
│ ::ffff:92.248.134.86 │
│ ::ffff:91.124.115.91 │
│ ::ffff:178.126.222.105 │
│ ::ffff:37.229.239.173 │
└────────────────────────┘
without subqueries
(I will not show you an example.
It should work as in any mature SQL DBMS.)
Drawbacks:
— only equi-JOIN;
— no condition pushdown to the right hand side;
(work in progress)
To extend ORDER BY clause in your favourite
Replacing/Aggregating/Collapsing MergeTrees
A column must be added at the same time:
ALTER TABLE t
ADD COLUMN z UInt64
MODIFY ORDER BY (x, y, z)
for distributed queries
(contributed by Spotify)
SET optimize_skip_unused_shards = 1
<merge_tree>
<use_minimalistic_part_header_in_zookeeper>
1
</use_minimalistic_part_header_in_zookeeper>
</merge_tree>
Not enabled by default.
Support for the three latest major releases:
19.4.2.7 19.3.7 19.1.14
Publicly available per commit builds:
Public Test Datasets
Automated per-commit performance tests
Added UBSan and TSan (along with ASan).
Adaptive index granularity
Row-level security
Column and table TTL expressions
Indexing by z-Order curve
DDL queries for dictionaries
ASOF JOIN (by Citadel Securities)
S3 import/export
Multiple storage volumes
Role Based Access Control
Aggregate functions for machine learning
Merge JOIN
Workload management
Optimization of ORDER BY and GROUP BY with table's order key.
Web site: https://clickhouse.com/
Google groups: https://groups.google.com/forum/#!forum/clickhouse
Maillist: [email protected]
Telegram chat: https://telegram.me/clickhouse_en
GitHub: https://github.com/ClickHouse/ClickHouse/
Twitter: https://twitter.com/ClickHouseDB