gitlab-org--gitlab-foss/doc/development/iterating_tables_in_batches.md

21 KiB

stage group info
Data Stores Database To determine the technical writer assigned to the Stage/Group associated with this page, see https://about.gitlab.com/handbook/engineering/ux/technical-writing/#assignments

Iterating tables in batches

Rails provides a method called in_batches that can be used to iterate over rows in batches. For example:

User.in_batches(of: 10) do |relation|
  relation.update_all(updated_at: Time.now)
end

Unfortunately, this method is implemented in a way that is not very efficient, both query and memory usage wise.

To work around this you can include the EachBatch module into your models, then use the each_batch class method. For example:

class User < ActiveRecord::Base
  include EachBatch
end

User.each_batch(of: 10) do |relation|
  relation.update_all(updated_at: Time.now)
end

This produces queries such as:

User Load (0.7ms)  SELECT  "users"."id" FROM "users" WHERE ("users"."id" >= 41654)  ORDER BY "users"."id" ASC LIMIT 1 OFFSET 1000
  (0.7ms)  SELECT COUNT(*) FROM "users" WHERE ("users"."id" >= 41654) AND ("users"."id" < 42687)

The API of this method is similar to in_batches, though it doesn't support all of the arguments that in_batches supports. You should always use each_batch unless you have a specific need for in_batches.

Avoid iterating over non-unique columns

One should proceed with extra caution, and possibly avoid iterating over a column that can contain duplicate values. When you iterate over an attribute that is not unique, even with the applied max batch size, there is no guarantee that the resulting batches do not surpass it. The following snippet demonstrates this situation when one attempt to select Ci::Build entries for users with id between 1 and 10,000, the database returns 1 215 178 matching rows.

[ gstg ] production> Ci::Build.where(user_id: (1..10_000)).size
=> 1215178

This happens because built relation is translated into the following query

[ gstg ] production> puts Ci::Build.where(user_id: (1..10_000)).to_sql
SELECT "ci_builds".* FROM "ci_builds" WHERE "ci_builds"."type" = 'Ci::Build' AND "ci_builds"."user_id" BETWEEN 1 AND 10000
=> nil

And queries which filter non-unique column by range WHERE "ci_builds"."user_id" BETWEEN ? AND ?, even though the range size is limited to a certain threshold (10,000 in the previous example) this threshold does not translate to the size of the returned dataset. That happens because when taking n possible values of attributes, one can't tell for sure that the number of records that contains them is less than n.

Column definition

EachBatch uses the primary key of the model by default for the iteration. This works most of the cases, however in some cases, you might want to use a different column for the iteration.

Project.distinct.each_batch(column: :creator_id, of: 10) do |relation|
  puts User.where(id: relation.select(:creator_id)).map(&:id)
end

The query above iterates over the project creators and prints them out without duplications.

NOTE: In case the column is not unique (no unique index definition), calling the distinct method on the relation is necessary. Using not unique column without distinct may result in each_batch falling into an endless loop as described in following issue.

EachBatch in data migrations

When dealing with data migrations the preferred way to iterate over a large volume of data is using EachBatch.

A special case of data migration is a background migration where the actual data modification is executed in a background job. The migration code that determines the data ranges (slices) and schedules the background jobs uses each_batch.

Efficient usage of each_batch

EachBatch helps to iterate over large tables. It's important to highlight that EachBatch does not magically solve all iteration-related performance problems, and it might not help at all in some scenarios. From the database point of view, correctly configured database indexes are also necessary to make EachBatch perform well.

Example 1: Simple iteration

Let's consider that we want to iterate over the users table and print the User records to the standard output. The users table contains millions of records, thus running one query to fetch the users likely times out.

Users table overview

This is a simplified version of the users table which contains several rows. We have a few smaller gaps in the id column to make the example a bit more realistic (a few records were already deleted). Currently, we have one index on the id field.

Loading all users into memory (avoid):

users = User.all

users.each { |user| puts user.inspect }

Use each_batch:

# Note: for this example I picked 5 as the batch size, the default is 1_000
User.each_batch(of: 5) do |relation|
  relation.each { |user| puts user.inspect }
end

How each_batch works

As the first step, it finds the lowest id (start id) in the table by executing the following database query:

SELECT "users"."id" FROM "users" ORDER BY "users"."id" ASC LIMIT 1

Reading the start ID value

Notice that the query only reads data from the index (INDEX ONLY SCAN), the table is not accessed. Database indexes are sorted so taking out the first item is a very cheap operation.

The next step is to find the next id (end id) which should respect the batch size configuration. In this example we used a batch size of 5. EachBatch uses the OFFSET clause to get a "shifted" id value.

SELECT "users"."id" FROM "users" WHERE "users"."id" >= 1 ORDER BY "users"."id" ASC LIMIT 1 OFFSET 5

Reading the end ID value

Again, the query only looks into the index. The OFFSET 5 takes out the sixth id value: this query reads a maximum of six items from the index regardless of the table size or the iteration count.

At this point, we know the id range for the first batch. Now it's time to construct the query for the relation block.

SELECT "users".* FROM "users" WHERE "users"."id" >= 1 AND "users"."id" < 302

Reading the rows from the users table

Notice the < sign. Previously six items were read from the index and in this query, the last value is "excluded". The query looks at the index to get the location of the five user rows on the disk and read the rows from the table. The returned array is processed in Ruby.

The first iteration is done. For the next iteration, the last id value is reused from the previous iteration in order to find out the next end id value.

SELECT "users"."id" FROM "users" WHERE "users"."id" >= 302 ORDER BY "users"."id" ASC LIMIT 1 OFFSET 5

Reading the second end ID value

Now we can easily construct the users query for the second iteration.

SELECT "users".* FROM "users" WHERE "users"."id" >= 302 AND "users"."id" < 353

Reading the rows for the second iteration from the users table

Example 2: Iteration with filters

Building on top of the previous example, we want to print users with zero sign-in count. We keep track of the number of sign-ins in the sign_in_count column so we write the following code:

users = User.where(sign_in_count: 0)

users.each_batch(of: 5) do |relation|
  relation.each { |user| puts user.inspect }
end

each_batch produces the following SQL query for the start id value:

SELECT "users"."id" FROM "users" WHERE "users"."sign_in_count" = 0 ORDER BY "users"."id" ASC LIMIT 1

Selecting only the id column and ordering by id forces the database to use the index on the id (primary key index) column however, we also have an extra condition on the sign_in_count column. The column is not part of the index, so the database needs to look into the actual table to find the first matching row.

Reading the index with extra filter

NOTE: The number of scanned rows depends on the data distribution in the table.

  • Best case scenario: the first user was never logged in. The database reads only one row.
  • Worst case scenario: all users were logged in at least once. The database reads all rows.

In this particular example, the database had to read 10 rows (regardless of our batch size setting) to determine the first id value. In a "real-world" application it's hard to predict whether the filtering causes problems or not. In the case of GitLab, verifying the data on a production replica is a good start, but keep in mind that data distribution on GitLab.com can be different from self-managed instances.

Improve filtering with each_batch

Specialized conditional index
CREATE INDEX index_on_users_never_logged_in ON users (id) WHERE sign_in_count = 0

This is how our table and the newly created index looks like:

Reading the specialized index

This index definition covers the conditions on the id and sign_in_count columns thus makes the each_batch queries very effective (similar to the simple iteration example).

It's rare when a user was never signed in so we a anticipate small index size. Including only the id in the index definition also helps to keep the index size small.

Index on columns

Later on, we might want to iterate over the table filtering for different sign_in_count values, in those cases we cannot use the previously suggested conditional index because the WHERE condition does not match with our new filter (sign_in_count > 10).

To address this problem, we have two options:

  • Create another, conditional index to cover the new query.
  • Replace the index with a more generalized configuration.

NOTE: Having multiple indexes on the same table and on the same columns could be a performance bottleneck when writing data.

Let's consider the following index (avoid):

CREATE INDEX index_on_users_never_logged_in ON users (id, sign_in_count)

The index definition starts with the id column which makes the index very inefficient from data selectivity point of view.

SELECT "users"."id" FROM "users" WHERE "users"."sign_in_count" = 0 ORDER BY "users"."id" ASC LIMIT 1

Executing the query above results in an INDEX ONLY SCAN. However, the query still needs to iterate over an unknown number of entries in the index, and then find the first item where the sign_in_count is 0.

Reading an ineffective index

We can improve the query significantly by swapping the columns in the index definition (prefer).

CREATE INDEX index_on_users_never_logged_in ON users (sign_in_count, id)

Reading a good index

The following index definition does not work well with each_batch (avoid).

CREATE INDEX index_on_users_never_logged_in ON users (sign_in_count)

Since each_batch builds range queries based on the id column, this index cannot be used efficiently. The DB reads the rows from the table or uses a bitmap search where the primary key index is also read.

"Slow" iteration

Slow iteration means that we use a good index configuration to iterate over the table and apply filtering on the yielded relation.

User.each_batch(of: 5) do |relation|
  relation.where(sign_in_count: 0).each { |user| puts user inspect }
end

The iteration uses the primary key index (on the id column) which makes it safe from statement timeouts. The filter (sign_in_count: 0) is applied on the relation where the id is already constrained (range). The number of rows is limited.

Slow iteration generally takes more time to finish. The iteration count is higher and one iteration could yield fewer records than the batch size. Iterations may even yield 0 records. This is not an optimal solution; however, in some cases (especially when dealing with large tables) this is the only viable option.

Using Subqueries

Using subqueries in your each_batch query does not work well in most cases. Consider the following example:

projects = Project.where(creator_id: Issue.where(confidential: true).select(:author_id))

projects.each_batch do |relation|
  # do something
end

The iteration uses the id column of the projects table. The batching does not affect the subquery. This means for each iteration, the subquery is executed by the database. This adds a constant "load" on the query which often ends up in statement timeouts. We have an unknown number of confidential issues, the execution time and the accessed database rows depend on the data distribution in the issues table.

NOTE: Using subqueries works only when the subquery returns a small number of rows.

Improving Subqueries

When dealing with subqueries, a slow iteration approach could work: the filter on creator_id can be part of the generated relation object.

projects = Project.all

projects.each_batch do |relation|
  relation.where(creator_id: Issue.where(confidential: true).select(:author_id))
end

If the query on the issues table itself is not performant enough, a nested loop could be constructed. Try to avoid it when possible.

projects = Project.all

projects.each_batch do |relation|
  issues = Issue.where(confidential: true)

  issues.each_batch do |issues_relation|
    relation.where(creator_id: issues_relation.select(:author_id))
  end
end

If we know that the issues table has many more rows than projects, it would make sense to flip the queries, where the issues table is batched first.

Using JOIN and EXISTS

When to use JOINS:

  • When there's a 1:1 or 1:N relationship between the tables where we know that the joined record (almost) always exists. This works well for "extension-like" tables:
    • projects - project_settings
    • users - user_details
    • users - user_statuses
  • LEFT JOIN works well in this case. Conditions on the joined table need to go to the yielded relation so the iteration is not affected by the data distribution in the joined table.

Example:

users = User.joins("LEFT JOIN personal_access_tokens on personal_access_tokens.user_id = users.id")

users.each_batch do |relation|
  relation.where("personal_access_tokens.name = 'name'")
end

EXISTS queries should be added only to the inner relation of the each_batch query:

User.each_batch do |relation|
  relation.where("EXISTS (SELECT 1 FROM ...")
end

Complex queries on the relation object

When the relation object has several extra conditions, the execution plans might become "unstable".

Example:

Issue.each_batch do |relation|
  relation
    .joins(:metrics)
    .joins(:merge_requests_closing_issues)
    .where("id IN (SELECT ...)")
    .where(confidential: true)
end

Here, we expect that the relation query reads the BATCH_SIZE of user records and then filters down the results according to the provided queries. The planner might decide that using a bitmap index lookup with the index on the confidential column is a better way to execute the query. This can cause an unexpectedly high amount of rows to be read and the query could time out.

Problem: we know for sure that the relation is returning maximum BATCH_SIZE of records however, the planner does not know this.

Common table expression (CTE) trick to force the range query to execute first:

Issue.each_batch(of: 1000) do |relation|
  cte = Gitlab::SQL::CTE.new(:batched_relation, relation.limit(1000))

  scope = cte
    .apply_to(Issue.all)
    .joins(:metrics)
    .joins(:merge_requests_closing_issues)
    .where("id IN (SELECT ...)")
    .where(confidential: true)

  puts scope.to_a
end

EachBatch vs BatchCount

When adding new counters for Service Ping, the preferred way to count records is using the Gitlab::Database::BatchCount class. The iteration logic implemented in BatchCount has similar performance characteristics like EachBatch. Most of the tips and suggestions for improving BatchCount mentioned above applies to BatchCount as well.

Iterate with keyset pagination

There are a few special cases where iterating with EachBatch does not work. EachBatch requires one distinct column (usually the primary key), which makes the iteration impossible for timestamp columns and tables with composite primary keys.

Where EachBatch does not work, you can use keyset pagination to iterate over the table or a range of rows. The scaling and performance characteristics are very similar to EachBatch.

Examples:

  • Iterate over the table in a specific order (timestamp columns) in combination with a tie-breaker if column user to sort by does not contain unique values.
  • Iterate over the table with composite primary keys.

Iterate over the issues in a project by creation date

You can use keyset pagination to iterate over any database column in a specific order (for example, created_at DESC). To ensure consistent order of the returned records with the same values for created_at, use a tie-breaker column with unique values (for example, id).

Assume you have the following index in the issues table:

idx_issues_on_project_id_and_created_at_and_id" btree (project_id, created_at, id)

Fetching records for further processing

The following snippet iterates over issue records within the project using the specified order (created_at, id).

scope = Issue.where(project_id: 278964).order(:created_at, :id) # id is the tie-breaker

iterator = Gitlab::Pagination::Keyset::Iterator.new(scope: scope)

iterator.each_batch(of: 100) do |records|
  puts records.map(&:id)
end

You can add extra filters to the query. This example only lists the issue IDs created in the last 30 days:

scope = Issue.where(project_id: 278964).where('created_at > ?', 30.days.ago).order(:created_at, :id) # id is the tie-breaker

iterator = Gitlab::Pagination::Keyset::Iterator.new(scope: scope)

iterator.each_batch(of: 100) do |records|
  puts records.map(&:id)
end

Updating records in the batch

For complex ActiveRecord queries, the .update_all method does not work well, because it generates an incorrect UPDATE statement. You can use raw SQL for updating records in batches:

scope = Issue.where(project_id: 278964).order(:created_at, :id) # id is the tie-breaker

iterator = Gitlab::Pagination::Keyset::Iterator.new(scope: scope)

iterator.each_batch(of: 100) do |records|
  ApplicationRecord.connection.execute("UPDATE issues SET updated_at=NOW() WHERE issues.id in (#{records.dup.reselect(:id).to_sql})")
end

NOTE: To keep the iteration stable and predictable, avoid updating the columns in the ORDER BY clause.

Iterate over the merge_request_diff_commits table

The merge_request_diff_commits table uses a composite primary key (merge_request_diff_id, relative_order), which makes EachBatch impossible to use efficiently.

To paginate over the merge_request_diff_commits table, you can use the following snippet:

# Custom order object configuration:
order = Gitlab::Pagination::Keyset::Order.build([
  Gitlab::Pagination::Keyset::ColumnOrderDefinition.new(
    attribute_name: 'merge_request_diff_id',
    order_expression: MergeRequestDiffCommit.arel_table[:merge_request_diff_id].asc,
    nullable: :not_nullable,
    distinct: false,
  ),
  Gitlab::Pagination::Keyset::ColumnOrderDefinition.new(
    attribute_name: 'relative_order',
    order_expression: MergeRequestDiffCommit.arel_table[:relative_order].asc,
    nullable: :not_nullable,
    distinct: false,
  )
])
MergeRequestDiffCommit.include(FromUnion) # keyset pagination generates UNION queries

scope = MergeRequestDiffCommit.order(order)

iterator = Gitlab::Pagination::Keyset::Iterator.new(scope: scope)

iterator.each_batch(of: 100) do |records|
  puts records.map { |record| [record.merge_request_diff_id, record.relative_order] }.inspect
end

Order object configuration

Keyset pagination works well with simple ActiveRecord order scopes (first example. However, in special cases, you need to describe the columns in the ORDER BY clause (second example) for the underlying keyset pagination library. When the ORDER BY configuration cannot be automatically determined by the keyset pagination library, an error is raised.

The code comments of the Gitlab::Pagination::Keyset::Order and Gitlab::Pagination::Keyset::ColumnOrderDefinition classes give an overview of the possible options for configuring the ORDER BY clause. You can also find a few code examples in the keyset pagination documentation.