A day in the life of a click house query

© 2022 Altinity, Inc.
A Day in the Life of
a ClickHouse Query
Intro to ClickHouse Internals
Robert Hodges & Altinity Engineering
1
10 February 2022

Let’s make some introductions
ClickHouse support and services including Altinity.Cloud
Authors of Altinity Kubernetes Operator for ClickHouse
and other open source projects
Robert Hodges
Database geek with 30+ years
on DBMS systems. Day job:
Altinity CEO
Altinity Engineering
Database geeks with centuries
of experience in DBMS and
applications
2

Foundations
3

Understands SQL
Runs on bare metal to cloud
Shared nothing architecture
Stores data in columns
Parallel and vectorized execution
Scales to many petabytes
Is Open source (Apache 2.0)
ClickHouse is a SQL Data Warehouse
It’s the core engine for
real-time analytics
ClickHouse
Event
Streams
ELT
Object
Storage
Interactive
Graphics
Dashboards
APIs
4

If you understand the engine you can make it faster
ClickHouse has a simple execution model–there’s no magic
Any developer can understand how it works
Knowledge leads to faster and more efﬁcient queries
5
(Another fast
engine!)

What happens
when you insert
data?
6

Let’s create a table!
CREATE TABLE IF NOT EXISTS sdata (
DevId Int32,
Type String,
MDate Date,
MDatetime DateTime,
Value Float64
) ENGINE = MergeTree()
PARTITION BY toYYYYMM(MDate)
ORDER BY (DevId, MDatetime)
7
Table engine type
How to break data into
parts
How to index and sort
data in each part
Table columns

Let’s now insert some data…
INSERT INTO sdata VALUES
(15, 'TEMP', '2018-01-01', '2018-01-01 23:29:55', 18.0),
(15, 'TEMP', '2018-01-01', '2018-01-01 23:30:56', 18.7)
8
(This is an example. Most people don’t
insert data this way!)

Part in Storage
How does ClickHouse process an insert?
9
INSERT INTO sdata
VALUES
(15, 'TEMP', . . .),
(15, 'TEMP', . . .)
2 rows in set. Elapsed: 0.271 sec.
ClickHouse Server
Parse/Plan
Respond
Load
Part in RAM
Sort rows ( table ORDER BY )

Part in Storage
How can we make this more efﬁcient? Parallelize!
10
set max_insert_threads=4
insert into ontime_test
select * from ontime
where toYear(FlightDate)
between 2000 and 2001
2 rows in set. Elapsed: 0.271 sec.
ClickHouse Server
Parse/Plan
Respond
Load
Parts in RAM

Parallelism affects speed and memory usage
11
insert into ontime_test
select * from ontime
where toYear(FlightDate)
between 2000 and 2001
. . .
. . .

OK, where did those awesome stats come from?
SELECT
event_time,
type,
is_initial_query,
query_duration_ms / 1000 AS duration,
read_rows,
read_bytes,
result_rows,
formatReadableSize(memory_usage) AS memory,
query
FROM system.query_log
WHERE (user = 'default') AND (type = 'QueryFinish')
ORDER BY event_time DESC
LIMIT 50
12

What’s going on down there when you INSERT?
Table
Part
Index Columns
Sparse index ﬁnds rows by Carrier,
Origin, FlightDate
Columns sorted
by Carrier, Origin,
FlightDate
Rows in the part
all belong to
same Year
Part
Index Columns
Part 13

Understanding what’s in a MergeTree part
/var/lib/clickhouse/data/airline/ontime
2017-07-01 AA
2017-07-01 EV
2017-07-01 UA
2017-07-02 AA
...
primary.idx
|
|
|
|
.mrk .bin
20170701_20170731_355_355_2/
(FlightDate, Carrier...) ActualElapsedTime Airline AirlineID...
|
|
|
|
.mrk .bin
|
|
|
|
.mrk .bin
Granule
Compressed
Block
Mark
14

Why MergeTree? Because it merges!
Part
Index Columns
Part
Index Columns
Rewritten, Bigger Part
Index Columns
15
Update and delete also rewrite parts

Bigger parts are more efﬁcient!
● Pick a PARTITION BY that gives nice, fat partitions ( 1-300GB, < 1000
total parts per table)
○ Can’t decide? Partition by month.
● Insert large blocks of data to avoid lots of merges afterwards
○ ClickHouse is ﬁne with tens of millions of rows!
● The simplest way to make blocks bigger is to batch input data
○ Avoid different partition keys in the same block
○ ClickHouse has parameters like max_insert_block_size but defaults are OK
○ Look at logs and actual part sizes to see if you need to do more
16

How can I see how big table parts are?
SELECT
table, partition, name,
marks, rows, data_compressed_bytes,
data_uncompressed_bytes, bytes_on_disk
FROM system.parts
WHERE active
AND level=0
AND database = 'default'
AND table = 'ontime_test'
ORDER BY table DESC, partition ASC, name ASC
17
Part is in use
Part has not been merged

Tips to optimized INSERT
Making INSERT faster
● Increase max_insert_threads (parallel creation of parts)
● Enable input_format_parallel_parsing to parallelize input parsing
○ Works for TSV/CSV/Values data
● Write bigger blocks (less merging afterwards)
Making INSERT less memory intensive
● Decrease max_insert_threads (reduces parts simultaneously in memory)
● Disable input_format_parallel_parsing
● Write smaller blocks (less memory required at INSERT time)
18

How do basic
queries work?
19

Aggregation is a key feature of analytic queries
20
SELECT Carrier,
avg(DepDelay) AS Delay
FROM ontime
GROUP BY Carrier
ORDER BY Delay DESC
Aggregates group measurements for one more more dimensions
Aggregate
Function
Dimension

How does ClickHouse process a query with aggregates?
21
SELECT Carrier,
avg(DepDelay)AS Delay
FROM ontime
GROUP BY Carrier
ORDER BY Delay DESC
┌─FlightDate─┬──────────────Delay─┐
│ 2016-12-17 │ 53.72058516196447 │
│ 1990-12-21 │ 43.87767499654839 │
│ 1990-12-22 │ 43.64346952962076 │
. . .
ClickHouse Server
Parse/Plan
Merge/Sort
Scan
In-RAM
Hash
Tables
Parts in Storage

How can you compute an average in parallel?
22
= 2
Numerator=6
Denominator=3
Scan
Merge
1 2 3 1 3 5 0 5 0 0
Numerator=9
Denominator=3
Numerator=5
Denominator=4
6 + 9 + 5
3 + 3 + 4 Partial
Aggregate

How does a ClickHouse thread do aggregation?
23
Merge/Sort
Scan Thread
Parts in Storage
AL => 4259/1070,
2385/415, …
DL => 20663/1198,
25166/2711, …
… Scan Thread
Hash Table
Other
Scan
Thread
Hash
Tables
Result
GROUP BY
Key
Partial
Aggregates

We can now understand aggregation performance drivers
24
SELECT Carrier,
avg(DepDelay)AS Delay
FROM ontime
GROUP BY Carrier
ORDER BY Delay DESC
LIMIT 50
Simple aggregate, short
GROUP BY key with few values
SELECT Carrier, FlightDate,
avg(DepDelay) AS Delay,
uniqExact(TailNum) AS Aircraft
FROM ontime
GROUP BY Carrier, FlightDate
ORDER BY Delay DESC
LIMIT 50
More complex aggregates, longer
GROUP BY with more values
3.4 sec
2.4 GB RAM
0.84 sec
1.6 KB RAM

Parallelism affects speed and memory usage
25
SELECT Origin, FlightDate,
avg(DepDelay) AS Delay,
uniqExact(TailNum) AS Aircraft
FROM ontime
WHERE Carrier='WN'
GROUP BY Origin, FlightDate
ORDER BY Delay DESC
LIMIT 5
SET max_threads = 1
. . .
SET max_threads = 4
. . .
SET max_threads = 16

This is a good time to mention ClickHouse memory limits
26
SQL
Query
max_memory_usage
(Default=10Gb)
Single query limit
SQL
Query
max_memory_usage_for_user
(Default=Unlimited)
All queries for a user
SQL
Query
SQL
Query
SQL
Query
All memory on server
ClickHouse
Process
max_server _memory_usage
(Default=90% of available RAM)

Tips to make aggregation queries faster
● Remove/exchange “heavy” aggregation functions
● Reduce the number of values in GROUP BY
● Increase max_threads (parallelism)
● Reduce I/O
○ Filter out unnecessary rows
○ Improve compression of data in storage
27

Tips to reduce memory usage in aggregation queries
● Remove/exchange “heavy” aggregation functions
● Reduce number of values in GROUP BY
● Change max_threads value
● Dump aggregates to external storage
○ SET max_bytes_before_external_group_by > 0
● Filter out unnecessary rows
28

How do joins
work?
29

JOIN combines data between tables
SELECT o.Dest,
any(a.Name) AS AirportName,
count(Dest) AS Flights
FROM ontime o JOIN airports a
ON a.IATA = o.Dest
GROUP BY Dest
ORDER BY Flights
DESC LIMIT 10
30
Join condition
“Left”
Table
“Right”
Table

How does ClickHouse process a query with a join?
Left Side
Table
(Big)
Merged
Result
Filter
Right Side
Table
(Small)
In-RAM
Hash
Table
Load the right side table
Scan left side table in
parallel, computing
aggregates and adding
joined data
Merge and sort results
31
Keys and
column
values!

Let’s look more deeply at what’s happening in the scan
. . .
IATA
. . .
airports
. . .
Dest
. . .
ontime
SELECT . . . FROM ontime o JOIN airports a ON a.IATA = o.Dest
32
S
c
a
n
ATL 576
1501
3302
…
Hartsfield Jackson Atlanta International Airport
…
ORD 255 Chicago O'Hare International Airport
Partial
aggregates

It would be more efﬁcient to join after aggregating
Left Side
Table
(Big)
Merge
Filter
Right Side
Table
(Small)
In-RAM
Hash
Table
Load the right side table
Scan left side table in
parallel, computing
aggregates
Merge, then join and sort
33
Join

You can do exactly that with a subquery
SELECT o.Dest, any(a.Name) AS AirportName,
count(Dest) AS Flights
FROM ontime o
JOIN default.airports a ON a.IATA = o.Dest
GROUP BY Dest ORDER BY Flights
DESC LIMIT 10
SELECT o.Dest, a.Name AS AirportName, o.Flights
FROM (
SELECT Dest, count(Dest) AS Flights
FROM ontime GROUP BY Dest ) AS o
JOIN default.airports a ON a.IATA = o.Dest
ORDER BY Flights DESC LIMIT 10
34
2.71 sec
19.9 MB RAM
0.663 sec
1.58 KB RAM

Simple ways to keep JOINs fast and efﬁcient
● Keep the right side table(s) overall size small
● Minimize the columns joined from the right side
● Add ﬁlter conditions to the right side table to reduce rows
● JOIN after aggregation if possible
● Use a Dictionary instead of a JOIN
○ Dictionaries are just loaded once and can be shared across queries
35
Pro tip: The SQL IN operator is also a join under the covers.

How does a
distributed query
work?
36

Example of a distributed data set with shards and replicas
clickhouse-0
ontime
_local
airports
ontime
clickhouse-1
ontime
_local
airports
ontime
clickhouse-2
ontime
_local
airports
ontime
clickhouse-3
ontime
_local
airports
ontime
Distributed
table
(No data)
Sharded,
replicated
table
(Partial data)
Fully
replicated
table
(All data)
37

Distributed send subqueries to multiple nodes
ontime
_local
ontime
Application
ontime
_local
ontime
ontime
_local
ontime
ontime
_local
ontime
Application
Innermost
subselect is
distributed
AggregateState
computed
locally
Aggregates
merged on
initiator node
38

Queries are pushed to all shards
SELECT Carrier, avg(DepDelay) AS Delay
FROM ontime
GROUP BY Carrier ORDER BY Delay DESC
SELECT Carrier, avg(DepDelay) AS Delay
FROM ontime_local
GROUP BY Carrier ORDER BY Delay DESC
39

ClickHouse pushes down JOINs by default
SELECT o.Dest d, a.Name n, count(*) c, avg(o.ArrDelayMinutes) ad
FROM default.ontime o
JOIN default.airports a ON (a.IATA = o.Dest)
GROUP BY d, n HAVING c > 100000 ORDER BY d DESC
LIMIT 10
SELECT Dest AS d, Name AS n, count() AS c, avg(ArrDelayMinutes) AS
ad
FROM default.ontime_local AS o
ALL INNER JOIN default.airports AS a ON a.IATA = o.Dest
GROUP BY d, n HAVING c > 100000 ORDER BY d DESC LIMIT 10
40

...Unless the left side “table” is a subquery
SELECT d, Name n, c AS flights, ad
FROM
(
SELECT Dest d, count(*) c, avg(ArrDelayMinutes) ad
FROM default.ontime
GROUP BY d HAVING c > 100000
ORDER BY ad DESC
) AS o
LEFT JOIN airports ON airports.IATA = o.d
LIMIT 10
Remote
Servers
41

It’s more complex when multiple tables are distributed
select foo from T1 where a in (select a from T2)
distributed_product_mode=?
local
select foo
from T1_local
where a in (
select a
from T2_local)
allow
select foo
from T1_local
where a in (
select a
from T2)
global
create temporary table
tmp Engine = Set
AS select a from T2;
select foo from
T1_local where a in
tmp;
(Subquery runs on
local table)
(Subquery runs on
distributed table) (Subquery runs on initiator;
broadcast to local temp table)
42

Tips to make distributed queries more efﬁcient
● Think about where your data are located
● Move WHERE and heavy grouping work to left hand side of join
● Use a subquery to order joins after the remote scan
● Use the query_log to see what actually executes on the remote node(s)
43

Where to learn
more
44

Where is the documentation?
ClickHouse ofﬁcial docs – https://clickhouse.com/docs/
Altinity Blog – https://altinity.com/blog/
Altinity Youtube Channel –
https://www.youtube.com/channel/UCE3Y2lDKl_ZfjaCrh62onYA
Altinity Knowledge Base – https://kb.altinity.com/
Meetups, other blogs, and external resources. Use your powers of Search!
45

References for this talk
Altinity Knowledge Base – https://kb.altinity.com/
ClickHouse Source Code – https://github.com/ClickHouse/ClickHouse
Talks and Blog Articles -
● ClickHouse Deep Dive, Alexey Milovidov
● Про JOIN’ы (в ClickHouse) - Artyem Zuikov
● Модификаторы DISTINCT и ORDER BY для всех агрегатных функций -
Soﬁa Sergeevna Borzenkova
● ClickHouse Kernel Analysis - Storage Structure and Query Acceleration of
MergeTree - Alibaba Cloud
46

Thank you!
Questions?
https://altinity.com
47

A day in the life of a click house query

More Related Content

What's hot (20)

Similar to A day in the life of a click house query (20)

Recently uploaded (20)

A day in the life of a click house query