Sidekiq worker attributes

Job urgency

Jobs can have an urgency attribute set, which can be :high, :low, or :throttled. These have the below targets:

To set a job's urgency, use the urgency class method:

class HighUrgencyWorker
  include ApplicationWorker

  urgency :high

  # ...
end

Latency sensitive jobs

If a large number of background jobs get scheduled at once, queueing of jobs may occur while jobs wait for a worker node to be become available. This is normal and gives the system resilience by allowing it to gracefully handle spikes in traffic. Some jobs, however, are more sensitive to latency than others.

In general, latency-sensitive jobs perform operations that a user could reasonably expect to happen synchronously, rather than asynchronously in a background worker. A common example is a write following an action. Examples of these jobs include:

1. A job which updates a merge request following a push to a branch.
2. A job which invalidates a cache of known branches for a project after a push to the branch.
3. A job which recalculates the groups and projects a user can see after a change in permissions.
4. A job which updates the status of a CI pipeline after a state change to a job in the pipeline.

When these jobs are delayed, the user may perceive the delay as a bug: for example, they may push a branch and then attempt to create a merge request for that branch, but be told in the UI that the branch does not exist. We deem these jobs to be urgency :high.

Extra effort is made to ensure that these jobs are started within a very short period of time after being scheduled. However, in order to ensure throughput, these jobs also have very strict execution duration requirements:

1. The median job execution time should be less than 1 second.
2. 99% of jobs should complete within 10 seconds.

If a worker cannot meet these expectations, then it cannot be treated as a urgency :high worker: consider redesigning the worker, or splitting the work between two different workers, one with urgency :high code that executes quickly, and the other with urgency :low, which has no execution latency requirements (but also has lower scheduling targets).

Changing a queue's urgency

On GitLab.com, we run Sidekiq in several shards, each of which represents a particular type of workload.

When changing a queue's urgency, or adding a new queue, we need to take into account the expected workload on the new shard. Note that, if we're changing an existing queue, there is also an effect on the old shard, but that always reduces work.

To do this, we want to calculate the expected increase in total execution time and RPS (throughput) for the new shard. We can get these values from:

• The Queue Detail dashboard has values for the queue itself. For a new queue, we can look for queues that have similar patterns or are scheduled in similar circumstances.
• The Shard Detail dashboard has Total Execution Time and Throughput (RPS). The Shard Utilization panel displays if there is currently any excess capacity for this shard.

We can then calculate the RPS * average runtime (estimated for new jobs) for the queue we're changing to see what the relative increase in RPS and execution time we expect for the new shard:

new_queue_consumption = queue_rps * queue_duration_avg
shard_consumption = shard_rps * shard_duration_avg

(new_queue_consumption / shard_consumption) * 100

If we expect an increase of less than 5%, then no further action is needed.

Otherwise, please ping @gitlab-org/scalability on the merge request and ask for a review.

Jobs with External Dependencies

Most background jobs in the GitLab application communicate with other GitLab services. For example, PostgreSQL, Redis, Gitaly, and Object Storage. These are considered to be "internal" dependencies for a job.

However, some jobs are dependent on external services in order to complete successfully. Some examples include:

1. Jobs which call web-hooks configured by a user.
2. Jobs which deploy an application to a k8s cluster configured by a user.

These jobs have "external dependencies". This is important for the operation of the background processing cluster in several ways:

1. Most external dependencies (such as web-hooks) do not provide SLOs, and therefore we cannot guarantee the execution latencies on these jobs. Since we cannot guarantee execution latency, we cannot ensure throughput and therefore, in high-traffic environments, we need to ensure that jobs with external dependencies are separated from high urgency jobs, to ensure throughput on those queues.
2. Errors in jobs with external dependencies have higher alerting thresholds as there is a likelihood that the cause of the error is external.

class ExternalDependencyWorker
  include ApplicationWorker

  # Declares that this worker depends on
  # third-party, external services in order
  # to complete successfully
  worker_has_external_dependencies!

  # ...
end

A job cannot be both high urgency and have external dependencies.

CPU-bound and Memory-bound Workers

Workers that are constrained by CPU or memory resource limitations should be annotated with the worker_resource_boundary method.

Most workers tend to spend most of their time blocked, waiting on network responses from other services such as Redis, PostgreSQL, and Gitaly. Since Sidekiq is a multi-threaded environment, these jobs can be scheduled with high concurrency.

Some workers, however, spend large amounts of time on-CPU running logic in Ruby. Ruby MRI does not support true multi-threading - it relies on the GIL to greatly simplify application development by only allowing one section of Ruby code in a process to run at a time, no matter how many cores the machine hosting the process has. For IO bound workers, this is not a problem, since most of the threads are blocked in underlying libraries (which are outside of the GIL).

If many threads are attempting to run Ruby code simultaneously, this leads to contention on the GIL which has the effect of slowing down all processes.

In high-traffic environments, knowing that a worker is CPU-bound allows us to run it on a different fleet with lower concurrency. This ensures optimal performance.

Likewise, if a worker uses large amounts of memory, we can run these on a bespoke low concurrency, high memory fleet.

Note that memory-bound workers create heavy GC workloads, with pauses of 10-50ms. This has an impact on the latency requirements for the worker. For this reason, memory bound, urgency :high jobs are not permitted and fail CI. In general, memory bound workers are discouraged, and alternative approaches to processing the work should be considered.

If a worker needs large amounts of both memory and CPU time, it should be marked as memory-bound, due to the above restriction on high urgency memory-bound workers.

Declaring a Job as CPU-bound

This example shows how to declare a job as being CPU-bound.

class CPUIntensiveWorker
  include ApplicationWorker

  # Declares that this worker will perform a lot of
  # calculations on-CPU.
  worker_resource_boundary :cpu

  # ...
end

Determining whether a worker is CPU-bound

We use the following approach to determine whether a worker is CPU-bound:

• In the Sidekiq structured JSON logs, aggregate the worker duration and cpu_s fields.
• duration refers to the total job execution duration, in seconds
• cpu_s is derived from the Process::CLOCK_THREAD_CPUTIME_ID counter, and is a measure of time spent by the job on-CPU.
• Divide cpu_s by duration to get the percentage time spend on-CPU.
• If this ratio exceeds 33%, the worker is considered CPU-bound and should be annotated as such.
• Note that these values should not be used over small sample sizes, but rather over fairly large aggregates.

Feature category

All Sidekiq workers must define a known feature category.

Job data consistency strategies

In GitLab 13.11 and earlier, Sidekiq workers would always send database queries to the primary database node, both for reads and writes. This ensured that data integrity is both guaranteed and immediate, since in a single-node scenario it is impossible to encounter stale reads even for workers that read their own writes. If a worker writes to the primary, but reads from a replica, however, the possibility of reading a stale record is non-zero due to replicas potentially lagging behind the primary.

When the number of jobs that rely on the database increases, ensuring immediate data consistency can put unsustainable load on the primary database server. We therefore added the ability to use Database Load Balancing for Sidekiq workers. By configuring a worker's data_consistency field, we can then allow the scheduler to target read replicas under several strategies outlined below.

Trading immediacy for reduced primary load

We require Sidekiq workers to make an explicit decision around whether they need to use the primary database node for all reads and writes, or whether reads can be served from replicas. This is enforced by a RuboCop rule, which ensures that the data_consistency field is set.

When setting this field, consider the following trade-off:

• Ensure immediately consistent reads, but increase load on the primary database.
• Prefer read replicas to add relief to the primary, but increase the likelihood of stale reads that have to be retried.

To maintain the same behavior compared to before this field was introduced, set it to :always, so database operations will only target the primary. Reasons for having to do so include workers that mostly or exclusively perform writes, or workers that read their own writes and who might run into data consistency issues should a stale record be read back from a replica. Try to avoid these scenarios, since :always should be considered the exception, not the rule.

To allow for reads to be served from replicas, we added two additional consistency modes: :sticky and :delayed.

When you declare either :sticky or :delayed consistency, workers become eligible for database load-balancing.

In both cases, if the replica is not up-to-date and the time from scheduling the job was less than the minimum delay interval, the jobs sleep up to the minimum delay interval (0.8 seconds). This gives the replication process time to finish. The difference is in what happens when there is still replication lag after the delay: sticky workers switch over to the primary right away, whereas delayed workers fail fast and are retried once. If they still encounter replication lag, they also switch to the primary instead. If your worker never performs any writes, it is strongly advised to apply one of these consistency settings, since it will never need to rely on the primary database node.

The table below shows the data_consistency attribute and its values, ordered by the degree to which they prefer read replicas and will wait for replicas to catch up:

In all cases workers read either from a replica that is fully caught up, or from the primary node, so data consistency is always ensured.

To set a data consistency for a worker, use the data_consistency class method:

class DelayedWorker
  include ApplicationWorker

  data_consistency :delayed

  # ...
end

feature_flag property

The feature_flag property allows you to toggle a job's data_consistency, which permits you to safely toggle load balancing capabilities for a specific job. When feature_flag is disabled, the job defaults to :always, which means that the job will always use the primary database.

The feature_flag property does not allow the use of feature gates based on actors. This means that the feature flag cannot be toggled only for particular projects, groups, or users, but instead, you can safely use percentage of time rollout. Note that since we check the feature flag on both Sidekiq client and server, rolling out a 10% of the time, will likely results in 1% (0.1 [from client]*0.1 [from server]) of effective jobs using replicas.

Example:

class DelayedWorker
  include ApplicationWorker

  data_consistency :delayed, feature_flag: :load_balancing_for_delayed_worker

  # ...
end

Data consistency with idempotent jobs

For idempotent jobs that declare either :sticky or :delayed data consistency, we are preserving the latest WAL location while deduplicating, ensuring that we read from the replica that is fully caught up.