debian-mirror-gitlab/doc/architecture/blueprints/ci_data_decay/pipeline_partitioning.md

stage	group	comments	description
none	unassigned	false	Pipeline data partitioning design

Pipeline data partitioning design

Disclaimer: The following contains information related to upcoming products, features, and functionality.

It is important to note that the information presented is for informational purposes only. Please do not rely on this information for purchasing or planning purposes.

As with all projects, the items mentioned in this document and linked pages are subject to change or delay. The development, release and timing of any products, features, or functionality remain at the sole discretion of GitLab Inc.

What problem are we trying to solve?

We want to partition the CI/CD dataset, because some of the database tables are extremely large, which might be challenging in terms of scaling single node reads, even after we ship the CI/CD database decomposition.

We want to reduce the risk of database performance degradation by transforming a few of the largest database tables into smaller ones using PostgreSQL declarative partitioning.

See more details about this effort in the parent blueprint.

How are CI/CD data decomposition, partitioning, and time-decay related?

CI/CD decomposition is an extraction of a CI/CD database cluster out of the "main" database cluster, to make it possible to have a different primary database receiving writes. The main benefit is doubling the capacity for writes and data storage. The new database cluster will not have to serve reads / writes for non-CI/CD database tables, so this offers some additional capacity for reads too.

CI/CD partitioning is dividing large CI/CD database tables into smaller ones. This will improve reads capacity on every CI/CD database node, because it is much less expensive to read data from small tables, than from large multi-terabytes tables. We can add more CI/CD database replicas to better handle the increase in the number of SQL queries that are reading data, but we need partitioning to perform a single read more efficiently. Performance in other aspects will improve too, because PostgreSQL will be more efficient in maintaining multiple small tables than in maintaining a very large database table.

CI/CD time-decay allows us to benefit from the strong time-decay characteristics of pipeline data. It can be implemented in many different ways, but using partitioning to implement time-decay might be especially beneficial. When implementing a time decay we usually mark data as archived, and migrate it out of a database to a different place when data is no longer relevant or needed. Our dataset is extremely large (tens of terabytes), so moving such a high volume of data is challenging. When time-decay is implemented using partitioning, we can archive the entire partition (or set of partitions) by simply updating a single record in one of our database tables. It is one of the least expensive ways to implement time-decay patterns at a database level.

Why do we need to partition CI/CD data?

We need to partition CI/CD data because our database tables storing pipelines, builds, and artifacts are too large. The ci_builds database table size is currently around 2.5 TB with an index of around 1.4 GB. This is too much and violates our principle of 100 GB max size. We also want to build alerting to notify us when this number is exceeded.

We’ve seen numerous S1 and S2 database-related production environment incidents, over the last couple of months, for example:

• S1: 2022-03-17 Increase in writes in ci_builds table
• S1: 2021-11-22 Excessive buffer read in replicas for ci_job_artifacts
• S2: 2022-04-12 Transactions detected that have been running for more than 10m
• S2: 2022-04-06 Database contention plausibly caused by excessive ci_builds reads
• S2: 2022-03-18 Unable to remove a foreign key on ci_builds

We have approximately 50 ci_* prefixed database tables, and some of them would benefit from partitioning.

A simple SQL query to get this data:

WITH tables AS (SELECT table_name FROM information_schema.tables WHERE table_name LIKE 'ci_%')
  SELECT table_name,
    pg_size_pretty(pg_total_relation_size(quote_ident(table_name))) AS total_size,
    pg_size_pretty(pg_relation_size(quote_ident(table_name))) AS table_size,
    pg_size_pretty(pg_indexes_size(quote_ident(table_name))) AS index_size,
    pg_total_relation_size(quote_ident(table_name)) AS total_size_bytes
  FROM tables ORDER BY total_size_bytes DESC;

See data from March 2022:

Table name	Total size	Index size
ci_builds	3.5 TB	1 TB
ci_builds_metadata	1.8 TB	150 GB
ci_job_artifacts	600 GB	300 GB
ci_pipelines	400 GB	300 GB
ci_stages	200 GB	120 GB
ci_pipeline_variables	100 GB	20 GB
(...around 40 more)

Based on the table above, it is clear that there are tables with a lot of stored data.

While we have almost 50 CI/CD-related database tables, we are initially interested in partitioning only 6 of them. We can start by partitioning the most interesting tables in an iterative way, but we also should have a strategy for partitioning the remaining ones if needed. This document is an attempt to capture this strategy, describe as many details as possible, to share this knowledge among engineering teams.

How do we want to partition CI/CD data?

We want to partition the CI/CD tables in iterations. It might not be feasible to partition all of the 6 initial tables at once, so an iterative strategy might be necessary. We also want to have a strategy for partitioning the remaining database tables when it becomes necessary.

It is also important to avoid large data migrations. We store almost 6 terabytes of data in the biggest CI/CD tables, in many different columns and indexes. Migrating this amount of data might be challenging and could cause instability in the production environment. Due to this concern, we’ve developed a way to attach an existing database table as a partition zero without downtime and excessive database locking, what has been demonstrated in one of the first proofs of concept. This makes creation of a partitioned schema possible without a downtime (for example using a routing table p_ci_pipelines), by attaching an existing ci_pipelines table as partition zero without exclusive locking. It will be possible to use the legacy table as usual, but we can create the next partition when needed and the p_ci_pipelines table will be used for routing queries. To use the routing table we need to find a good partitioning key.

Our plan is to use logical partition IDs. We want to start with the ci_pipelines table and create a partition_id column with a DEFAULT value of 100 or 1000. Using a DEFAULT value avoids the challenge of backfilling this value for every row. Adding a CHECK constraint prior to attaching the first partition tells PostgreSQL that we’ve already ensured consistency and there is no need to check it while holding an exclusive table lock when attaching this table as a partition to the routing table (partitioned schema definition). We will increment this value every time we create a new partition for p_ci_pipelines, and the partitioning strategy will be LIST partitioning.

We will also create a partition_id column in the other initial 6 database tables we want to iteratively partition. After a new pipeline is created, it will get a partition_id assigned, and all the related resources, like builds and artifacts, will share the same value. We want to add the partition_id column into all 6 problematic tables because we can avoid backfilling this data when we decide it is time to start partitioning them.

We want to partition CI/CD data iteratively, so we will start with the pipelines table, and create at least one, but likely two, partitions. The pipelines table will be partitioned using the LIST partitioning strategy. It is possible that, after some time, p_ci_pipelines will store data in two partitions with IDs of 100 and 101. Then we will try partitioning ci_builds. Therefore we might want to use RANGE partitioning in p_ci_builds with IDs 100 and 101, because builds for the two logical partitions used will still be stored in a single table.

Physical partitioning and logical partitioning will be separated, and a strategy will be determined when we implement partitioning for the respective database tables. Using RANGE partitioning works similarly to using LIST partitioning in database tables other than ci_pipelines, but because we can guarantee continuity of partition_id values, using RANGE partitioning might be a better strategy.

Why do we want to use explicit logical partition ids?

Partitioning CI/CD data using a logical partition_id has several benefits. We could partition by a primary key, but this would introduce much more complexity and additional cognitive load required to understand how the data is being structured and stored in partitions.

CI/CD data is hierarchical data. Stages belong to pipelines, builds belong to stages, artifacts belong to builds (with rare exceptions). We are designing a partitioning strategy that reflects this hierarchy, to reduce the complexity and therefore cognitive load for contributors. With an explicit partition_id associated with a pipeline, we can cascade the partition ID number when trying to retrieve all resources associated with a pipeline. We know that for a pipeline 12345 with a partition_id of 102, we are always able to find associated resources in logical partitions with number 102 in other routing tables, and PostgreSQL will know in which partitions these records are being stored in for every table.

Another interesting benefit for using a single and incremental latest partition_id number, associated with pipelines, is that in theory we can cache it in Redis or in memory to avoid excessive reads from the database to find this number, though we might not need to do this.

The single and uniform partition_id value for pipeline data gives us more choices later on than primary-keys-based partitioning.

Splitting large partitions into smaller ones

We want to start with the initial pipeline_id number 100 (or higher, like 1000, depending on our calculations and estimations). We do not want to start from 1, because existing tables are also large already, and we might want to split them into smaller partitions. If we start with 100, we will be able to create partitions for partition_id of 1, 20, 45, and move existing records there by updating partition_id from 100 to a smaller number.

PostgreSQL will move these records into their respective partitions in a consistent way, provided that we do it in a transaction for all pipeline resources at the same time. If we ever decide to split large partitions into smaller ones (it's not yet clear if we will need to do this), we might be able to just use background migrations to update partition IDs, and PostgreSQL is smart enough to move rows between partitions on its own.

Storing partitions metadata in the database

In order to build an efficient mechanism that will be responsible for creating new partitions, and to implement time decay we want to introduce a partitioning metadata table, called ci_partitions. In that table we would store metadata about all the logical partitions, with many pipelines per partition. We may need to store a range of pipeline ids per logical partition. Using it we will be able to find the partition_id number for a given pipeline ID and we will also find information about which logical partitions are “active” or “archived”, which will help us to implement a time-decay pattern using database declarative partitioning.

ci_partitions table will store information about a partition identifier, pipeline ids range it is valid for and whether the partitions have been archived or not. Additional columns with timestamps may be helpful too.

Implementing a time-decay pattern using partitioning

We can use ci_partitions to implement a time-decay pattern using declarative partitioning. By telling PostgreSQL which logical partitions are archived we can stop reading from these partitions using a SQL query like the one below.

SELECT * FROM ci_builds WHERE partition_id IN (
  SELECT id FROM ci_partitions WHERE active = true
);

This query will make it possible to limit the number of partitions we will read from, and therefore will cut access to "archived" pipeline data, using our data retention policy for CI/CD data. Ideally we do not want to read from more than two partitions at once, so we need to align the automatic partitioning mechanisms with the time-decay policy. We will still need to implement new access patterns for the archived data, presumably through the API, but the cost of storing archived data in PostgreSQL will be reduced significantly this way.

There are some technical details here that are out of the scope of this description, but by using this strategy we can "archive" data, and make it much less expensive to reside in our PostgreSQL cluster by simply toggling a boolean column value.

Accessing partitioned data

It will be possible to access partitioned data whether it has been archived or not, in most places in GitLab. On a merge request page, we will always show pipeline details even if the merge request was created years ago. We can do that because ci_partitions will be a lookup table associating a pipeline ID with its partition_id, and we will be able to find the partition that the pipeline data is stored in.

We will need to constrain access to searching through pipelines, builds, artifacts etc. Search can not be done through all partitions, as it would not be efficient enough, hence we will need to find a better way of searching through archived pipelines data. It will be necessary to have different access patterns to access archived data in the UI and API.

There are a few challenges in enforcing usage of the partition_id partitioning key in PostgreSQL. To make it easier to update our application to support this, we have designed a new queries analyzer in our proof of concept merge request. It helps to find queries that are not using the partitioning key.

In a separate proof of concept merge request and related issue we demonstrated that using the uniform partition_id makes it possible to extend Rails associations with an additional scope modifier so we can provide the partitioning key in the SQL query.

Using instance dependent associations, we can easily append a partitioning key to SQL queries that are supposed to retrieve associated pipeline resources, for example:

has_many :builds, -> (pipeline) { where(partition_id: pipeline.partition_id) }

The problem with this approach is that it makes preloading much more difficult as instance dependent associations can not be used with preloads:

ArgumentError: The association scope 'builds' is instance dependent (the
scope block takes an argument). Preloading instance dependent scopes is not
supported.

We also need to build a proof of concept for removing data on the PostgreSQL side (using foreign keys with ON DELETE CASCADE) and removing data through Rails associations, as this might be an important area of uncertainty.

We need to better understand how unique constraints we are currently using will perform when using the partitioned schema.

We have also designed a query analyzer that makes it possible to detect direct usage of zero partitions, legacy tables that have been attached as first partitions to routing tables, to ensure that all queries are targeting partitioned schema or partitioned routing tables, like p_ci_pipelines.

Why not partition using the project or namespace ID?

We do not want to partition using project_id or namespace_id because sharding and podding is a different problem to solve, on a different layer of the application. It doesn't solve the original problem statement of performance growing worse over time as we build up infrequently read data. We may want to introduce pods in the future, and that might become the primary mechanism of separating data based on the group or project the data is associated with.

In theory we could use either project_id or namespace_id as a second partitioning dimension, but this would add more complexity to a problem that is already very complex.

Partitioning builds queuing tables

We also want to partition our builds queuing tables. We currently have two: ci_pending_builds and ci_running_builds. These tables are different from other CI/CD data tables, as there are business rules in our product that make all data stored in them invalid after 24 hours.

As a result, we will need to use a different strategy to partition those database tables, by removing partitions entirely after these are older than 24 hours, and always reading from two partitions through a routing table. The strategy to partition these tables is well understood, but requires a solid Ruby-based automation to manage the creation and deletion of these partitions. To achieve that we will collaborate with the Database team to adapt existing database partitioning tools to support CI/CD data partitioning.

Iterating to reduce the risk

This strategy should reduce the risk of implementing CI/CD partitioning to acceptable levels. We are also focusing on implementing partitioning for reading only from two partitions initially to make it possible to detach zero partitions in case of problems in our production environment. Every iteration phase, described below has a revert strategy and before shipping database changes we want to test them in our benchmarking environment.

The main way of reducing risk in case of this effort is iteration and making things reversible. Shipping changes, described in this document, in a safe and reliable way is our priority.

As we move forward with the implementation we will need to find even more ways to iterate on the design, support incremental rollouts and have better control over reverting changes in case of something going wrong. It is sometimes challenging to ship database schema changes iteratively, and even more difficult to support incremental rollouts to the production environment. This can, however, be done, it just sometimes requires additional creativity, that we will certainly need here. Some examples of how this could look like:

Incremental rollout of partitioned schema

Once we introduce a first partitioned routing table (presumably p_ci_pipelines) and attach its zero partition (ci_pipelines), we will need to start interacting with the new routing table, instead of a concrete partition zero. Usually we would override the database table the Ci::Pipeline Rails model would use with something like self.table_name = 'p_ci_pipelines'. Unfortunately this approach might not support incremental rollout, because self.table_name will be read upon application boot up, and later we might be unable revert this change without restarting the application.

One way of solving this might be introducing Ci::Partitioned::Pipeline model, that will inherit from Ci::Pipeline. In that model we would set self.table_name to p_ci_pipeline and return its meta class from Ci::Pipeline.partitioned as a scope. This will allow us to use feature flags to route reads from ci_pipelines to p_ci_pipelines with a simple revert strategy.

Incremental experimentation with partitioned reads

Another example would be related to the time when we decide to attach another partition. The goal of Phase 1 will be have two partitions per partitioned schema / routing table, meaning that for p_ci_pipelines we will have ci_pipelines attached as partition zero, and a new ci_pipelines_p1 partition created for new data. All reads from p_ci_pipelines will also need to read data from the p1 partition and we should also iteratively experiment with reads targeting more than one partition, to evaluate performance and overhead of partitioning.

We can do that by moving old data to ci_pipelines_m1 (minus 1) partition iteratively. Perhaps we will create partition_id = 1 and move some really old pipelines there. We can then iteratively migrate data into m1 partition to measure the impact, performance and increase our confidence before creating a new partition p1 for new (still not created) data.

Iterations

We want to focus on Phase 1 iteration first. The goal and the main objective of this iteration is to partition the biggest 6 CI/CD database tables into 6 routing tables (partitioned schema) and 12 partitions. This will leave our Rails SQL queries mostly unchanged, but it will also make it possible to perform emergency detachment of "zero partitions" if there is a database performance degradation. This will cut users off their old data, but the application will remain up and running, which is a better alternative to application-wide outage.

1. Phase 0: Build CI/CD data partitioning strategy: Done. ✅

2. Phase 1: Partition the 6 biggest CI/CD database tables.

   1. Create partitioned schemas for all 6 database tables.
   2. Design a way to cascade partition_id to all partitioned resources.
   3. Implement initial query analyzers validating that we target routing tables.
   4. Attach zero partitions to the partitioned database tables.
   5. Update the application to target routing tables and partitioned tables.
   6. Measure the performance and efficiency of this solution.

   Revert strategy: Switch back to using concrete partitions instead of routing tables.

3. Phase 2: Add a partitioning key to add SQL queries targeting partitioned tables.

   1. Implement query analyzer to check if queries targeting partitioned tables are using proper partitioning keys.
   2. Modify existing queries to make sure that all of them are using a partitioning key as a filter.

   Revert strategy: Use feature flags, query by query.

4. Phase 3: Build new partitioned data access patterns.

   1. Build a new API or extend an existing one to allow access to data stored in partitions that are supposed to be excluded based on the time-decay data retention policy.

   Revert strategy: Feature flags.

5. Phase 4: Introduce time-decay mechanisms built on top of partitioning.

   1. Build time-decay policy mechanisms.
   2. Enable the time-decay strategy on GitLab.com.

6. Phase 5: Introduce mechanisms for creating partitions automatically.

   1. Make it possible to create partitions in an automatic way.
   2. Deliver the new architecture to self-managed instances.

Conclusions

We want to build a solid strategy for partitioning CI/CD data. We are aware of the fact that it is difficult to iterate on this design, because a mistake made in managing the database schema of our multi-terabyte PostgreSQL instance might not be easily reversible without potential downtime. That is the reason we are spending a significant amount of time to research and refine our partitioning strategy. The strategy, described in this document, is subject to iteration as well. Whenever we find a better way to reduce the risk and improve our plan, we should update this document as well.

We’ve managed to find a way to avoid large-scale data migrations, and we are building an iterative strategy for partitioning CI/CD data. We documented our strategy here to share knowledge and solicit feedback from other team members.

Who

Authors:

Role	Who
Author	Grzegorz Bizon

Recommenders:

Role	Who
Distingiushed Engineer	Kamil Trzciński