The engine inherits from [MergeTree](../../../engines/table-engines/mergetree-family/mergetree.md) and adds the logic of rows collapsing to data parts merge algorithm.
`CollapsingMergeTree` asynchronously deletes (collapses) pairs of rows if all of the fields in a sorting key (`ORDER BY`) are equivalent excepting the particular field `Sign` which can have `1` and `-1` values. Rows without a pair are kept. For more details see the [Collapsing](#table_engine-collapsingmergetree-collapsing) section of the document.
When creating a `CollapsingMergeTree` table, the same [query clauses](../../../engines/table-engines/mergetree-family/mergetree.md#table_engine-mergetree-creating-a-table) are required, as when creating a `MergeTree` table.
Consider the situation where you need to save continually changing data for some object. It sounds logical to have one row for an object and update it at any change, but update operation is expensive and slow for DBMS because it requires rewriting of the data in the storage. If you need to write data quickly, update not acceptable, but you can write the changes of an object sequentially as follows.
Use the particular column `Sign`. If `Sign = 1` it means that the row is a state of an object, let’s call it “state” row. If `Sign = -1` it means the cancellation of the state of an object with the same attributes, let’s call it “cancel” row.
For example, we want to calculate how much pages users checked at some site and how long they were there. At some moment we write the following row with the state of user activity:
1. The program that writes the data should remember the state of an object to be able to cancel it. “Cancel” string should contain copies of the sorting key fields of the “state” string and the opposite `Sign`. It increases the initial size of storage but allows to write the data quickly.
2. Long growing arrays in columns reduce the efficiency of the engine due to load for writing. The more straightforward data, the higher the efficiency.
3. The `SELECT` results depend strongly on the consistency of object changes history. Be accurate when preparing data for inserting. You can get unpredictable results in inconsistent data, for example, negative values for non-negative metrics such as session depth.
When ClickHouse merges data parts, each group of consecutive rows with the same sorting key (`ORDER BY`) is reduced to not more than two rows, one with `Sign = 1` (“state” row) and another with `Sign = -1` (“cancel” row). In other words, entries collapse.
Also when there are at least 2 more “state” rows than “cancel” rows, or at least 2 more “cancel” rows then “state” rows, the merge continues, but ClickHouse treats this situation as a logical error and records it in the server log. This error can occur if the same data were inserted more than once.
The `Sign` is required because the merging algorithm doesn’t guarantee that all of the rows with the same sorting key will be in the same resulting data part and even on the same physical server. ClickHouse process `SELECT` queries with multiple threads, and it can not predict the order of rows in the result. The aggregation is required if there is a need to get completely “collapsed” data from `CollapsingMergeTree` table.
To finalize collapsing, write a query with `GROUP BY` clause and aggregate functions that account for the sign. For example, to calculate quantity, use `sum(Sign)` instead of `count()`. To calculate the sum of something, use `sum(Sign * x)` instead of `sum(x)`, and so on, and also add `HAVING sum(Sign) > 0`.
The aggregates `count`, `sum` and `avg` could be calculated this way. The aggregate `uniq` could be calculated if an object has at least one state not collapsed. The aggregates `min` and `max` could not be calculated because `CollapsingMergeTree` does not save the values history of the collapsed states.
If you need to extract data without aggregation (for example, to check whether rows are present whose newest values match certain conditions), you can use the `FINAL` modifier for the `FROM` clause. This approach is significantly less efficient.
We use two `INSERT` queries to create two different data parts. If we insert the data with one query ClickHouse creates one data part and will not perform any merge ever.
With two `INSERT` queries, we created 2 data parts. The `SELECT` query was performed in 2 threads, and we got a random order of rows. Collapsing not occurred because there was no merge of the data parts yet. ClickHouse merges data part in an unknown moment which we can not predict.
The idea is that merges take into account only key fields. And in the “Cancel” line we can specify negative values that equalize the previous version of the row when summing without using the Sign column. For this approach, it is necessary to change the data type `PageViews`,`Duration` to store negative values of UInt8 -\> Int16.
