- Job urgency
- Jobs with External Dependencies
- CPU-bound and Memory-bound Workers
- Declaring a Job as CPU-bound
- Determining whether a worker is CPU-bound
- Feature category
- Job data consistency strategies
- Job pause control
Worker classes can define certain attributes to control their behavior and add metadata.
Child classes inheriting from other workers also inherit these attributes, so you only have to redefine them if you want to override their values.
Jobs can have an
urgency attribute set, which can be
:throttled. These have the below targets:
|Urgency||Queue Scheduling Target||Execution Latency Requirement|
|10 seconds||10 seconds|
||1 minute||5 minutes|
To set a job’s urgency, use the
urgency class method:
class HighUrgencyWorker include ApplicationWorker urgency :high # ... end
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 standard 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:
- A job which updates a merge request following a push to a branch.
- A job which invalidates a cache of known branches for a project after a push to the branch.
- A job which recalculates the groups and projects a user can see after a change in permissions.
- 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
Extra effort is made to ensure that these jobs are started within a very short period of time after being scheduled. However, to ensure throughput, these jobs also have very strict execution duration requirements:
- The median job execution time should be less than 1 second.
- 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).
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. 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.
@gitlab-org/scalability on the merge request and ask
for a review.
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 to complete successfully. Some examples include:
- Jobs which call web-hooks configured by a user.
- Jobs which deploy an application to a Kubernetes cluster configured by a user.
These jobs have “external dependencies”. This is important for the operation of the background processing cluster in several ways:
- 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.
- 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.
Workers that are constrained by CPU or memory resource limitations should be
annotated with the
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.
Memory-bound workers create heavy GC workloads, with pauses of
10-50 ms. This has an impact on the latency requirements for the
worker. For this reason,
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
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.
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
We use the following approach to determine whether a worker is CPU-bound:
- In the Sidekiq structured JSON logs, aggregate the worker
durationrefers to the total job execution duration, in seconds
cpu_sis derived from the
Process::CLOCK_THREAD_CPUTIME_IDcounter, and is a measure of time spent by the job on-CPU.
durationto get the percentage time spend on-CPU.
- If this ratio exceeds 33%, the worker is considered CPU-bound and should be annotated as such.
- These values should not be used over small sample sizes, but rather over fairly large aggregates.
All Sidekiq workers must define a known feature category.
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.
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
database operations 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:
:delayed. A RuboCop rule
reminds the developer when
:always data consistency mode is used. If workers require the primary database, you can disable the rule in-line.
When you declare either
:delayed consistency, workers become eligible for database
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:
switch over to the primary right away, whereas
delayed workers fail fast and are retried once.
If the workers still encounter replication lag, they switch to the primary instead. If your worker never performs any writes,
it is strongly advised to apply
:delayed consistency settings, since the worker never needs 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 wait for replicas to catch up:
|The job is required to use the primary database (default).||It should be used for workers that primarily perform writes, have strict requirements around data consistency when reading their own writes, or are cron jobs.|
|The job prefers replicas, but switches to the primary for writes or when encountering replication lag.||It should be used for jobs that require to be executed as fast as possible but can sustain a small initial queuing delay.|
|The job prefers replicas, but switches to the primary for writes. When encountering replication lag before the job starts, the job is retried once. If the replica is still not up to date on the next retry, it switches to the primary.||It should be used for jobs where delaying execution further typically does not matter, such as cache expiration or web hooks execution.|
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 allows you to toggle a job’s
which permits you to safely toggle load balancing capabilities for a specific job.
feature_flag is disabled, the job defaults to
:always, which means that the job always uses the primary database.
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.
Since we check the feature flag on both Sidekiq client and server, rolling out a 10% of the time,
likely results in 1% (
[from server]) of effective jobs using replicas.
class DelayedWorker include ApplicationWorker data_consistency :delayed, feature_flag: :load_balancing_for_delayed_worker # ... end
For idempotent jobs that declare either
:delayed data consistency, we are
preserving the latest WAL location while deduplicating,
ensuring that we read from the replica that is fully caught up.
pause_control property, you can conditionally pause job processing. If the strategy is active, the job
is stored in a separate
ZSET and re-enqueued when the strategy becomes inactive.
PauseControl::ResumeWorker is a cron
worker that checks if any paused jobs must be restarted.
pause_control, you can:
- Use one of the strategies defined in
- Define a custom strategy in
lib/gitlab/sidekiq_middleware/pause_control/strategies/and add the strategy to
module Gitlab module SidekiqMiddleware module PauseControl module Strategies class CustomStrategy < Base def should_pause? ApplicationSetting.current.elasticsearch_pause_indexing? end end end end end end
class PausedWorker include ApplicationWorker pause_control :custom_strategy # ... end