In this blog post, we’ll look at nested data structures in ClickHouse and how this can be used with PMM to look at queries.
Nested structures are not common in Relational Database Management Systems. Usually, it’s just flat tables. Sometimes it would be convenient to store unstructured information in structured databases.
We are working to adapt ClickHouse as a long term storage for Percona Monitoring and Management (PMM), and particularly to store detailed information about queries. One of the problems we are trying to solve is to count the different errors that cause a particular query to fail.
For example, for date 2017-08-17 the query:
|
1 |
"SELECT foo FROM bar WHERE id=?" |
was executed 1000 times. 25 times it failed with error code “1212”, and eight times it failed with error code “1250”. Of course, the traditional way to store this in relational data would be to have a table "Date, QueryID, ErrorCode, ErrorCnt" and then perform a JOIN to this table. Unfortunately, columnar databases don’t perform well with multiple joins, and often the recommendation is to have de-normalized tables.
We can create a column for each possible ErrorCode, but this is not an optimal solution. There could be thousands of them, and most of the time they would be empty.
In this case, ClickHouse proposes Nested data structures. For our case, these can be defined as:
|
1 |
CREATE TABLE queries<br>(<br> Period Date,<br> QueryID UInt32,<br> Fingerprint String,<br> Errors Nested<br> (<br> ErrorCode String,<br> ErrorCnt UInt32<br> )<br>)Engine=MergeTree(Period,QueryID,8192); |
This solution has obvious questions: How do we insert data into this table? How do we extract it?
Let’s start with INSERT. Insert can look like:
|
1 |
INSERT INTO queries VALUES ('2017-08-17',5,'SELECT foo FROM bar WHERE id=?',['1220','1230','1212'],[5,6,2])<br> |
which means that the inserted query during 2017-08-17 gave error 1220 five times, error 1230 six times and error 1212 two times.
Now, during a different date, it might produce different errors:
|
1 |
INSERT INTO queries VALUES ('2017-08-18',5,'SELECT foo FROM bar WHERE id=?',['1220','1240','1258'],[3,2,1])<br> |
Let’s take a look at ways to SELECT data. A very basic SELECT:
|
1 |
<br>SELECT *<br>FROM queries <br><br>┌─────Period─┬─QueryID─┬─Fingerprint─┬─Errors.ErrorCode───────┬─Errors.ErrorCnt─┐<br>│ 2017-08-17 │ 5 │ SELECT foo │ ['1220','1230','1212'] │ [5,6,2] │<br>│ 2017-08-18 │ 5 │ SELECT foo │ ['1220','1240','1260'] │ [3,16,12] │<br>└────────────┴─────────┴─────────────┴────────────────────────┴─────────────────┘<br> |
If we want to use a more familiar tabular output, we can use the ARRAY JOIN extension:
|
1 |
SELECT *<br>FROM queries <br>ARRAY JOIN Errors<br><br>┌─────Period─┬─QueryID─┬─Fingerprint─┬─Errors.ErrorCode─┬─Errors.ErrorCnt─┐<br>│ 2017-08-17 │ 5 │ SELECT foo │ 1220 │ 5 │<br>│ 2017-08-17 │ 5 │ SELECT foo │ 1230 │ 6 │<br>│ 2017-08-17 │ 5 │ SELECT foo │ 1212 │ 2 │<br>│ 2017-08-18 │ 5 │ SELECT foo │ 1220 │ 3 │<br>│ 2017-08-18 │ 5 │ SELECT foo │ 1240 │ 16 │<br>│ 2017-08-18 │ 5 │ SELECT foo │ 1260 │ 12 │<br>└────────────┴─────────┴─────────────┴──────────────────┴─────────────────┘ |
However, usually we want to see the aggregation over multiple periods, which can be done with traditional aggregation functions:
|
1 |
SELECT <br> QueryID,<br> Errors.ErrorCode,<br> SUM(Errors.ErrorCnt)<br>FROM queries<br>ARRAY JOIN Errors<br>GROUP BY <br> QueryID,<br> Errors.ErrorCode<br><br>┌─QueryID─┬─Errors.ErrorCode─┬─SUM(Errors.ErrorCnt)─┐<br>│ 5 │ 1212 │ 2 │<br>│ 5 │ 1230 │ 6 │<br>│ 5 │ 1260 │ 12 │<br>│ 5 │ 1240 │ 16 │<br>│ 5 │ 1220 │ 8 │<br>└─────────┴──────────────────┴──────────────────────┘ |
If we want to get really creative and return only one row per QueryID, we can do that as well:
|
1 |
SELECT <br> QueryID, <br> groupArray((ecode, cnt))<br>FROM <br>(<br> SELECT <br> QueryID, <br> ecode, <br> sum(ecnt) AS cnt<br> FROM queries <br> ARRAY JOIN <br> Errors.ErrorCode AS ecode, <br> Errors.ErrorCnt AS ecnt<br> GROUP BY <br> QueryID, <br> ecode<br>) <br>GROUP BY QueryID<br><br>┌─QueryID─┬─groupArray(tuple(ecode, cnt))──────────────────────────────┐<br>│ 5 │ [('1230',6),('1212',2),('1260',12),('1220',8),('1240',16)] │<br>└─────────┴────────────────────────────────────────────────────────────┘ |
ClickHouse provides flexible ways to store data in a less structured manner and variety of functions to extract and aggregate it – despite being a columnar database.
Happy data warehousing!
