9.9 KiB
stage | group | info |
---|---|---|
none | unassigned | 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 |
Generating chaos in a test GitLab instance
As Werner Vogels, the CTO at Amazon Web Services, famously put it, Everything fails, all the time.
As a developer, it's as important to consider the failure modes in which your software may operate as much as normal operation. Doing so can mean the difference between a minor hiccup leading to a scattering of 500
errors experienced by a tiny fraction of users, and a full site outage that affects all users for an extended period.
To paraphrase Tolstoy, all happy servers are alike, but all failing servers are failing in their own way. Luckily, there are ways we can attempt to simulate these failure modes, and the chaos endpoints are tools for assisting in this process.
Currently, there are four endpoints for simulating the following conditions:
- Slow requests.
- CPU-bound requests.
- Memory leaks.
- Unexpected process crashes.
Enabling chaos endpoints
For obvious reasons, these endpoints are not enabled by default on production
.
They are enabled by default on development environments.
WARNING: It is required that you secure access to the chaos endpoints using a secret token. You should not enable them in production unless you absolutely know what you're doing.
A secret token can be set through the GITLAB_CHAOS_SECRET
environment variable.
For example, when using the GDK
this can be done with the following command:
GITLAB_CHAOS_SECRET=secret gdk run
Replace secret
with your own secret token.
Invoking chaos
After you have enabled the chaos endpoints and restarted the application, you can start testing using the endpoints.
By default, when invoking a chaos endpoint, the web worker process which receives the request handles it. This means, for example, that if the Kill
operation is invoked, the Puma or Unicorn worker process handling the request is killed. To test these operations in Sidekiq, the async
parameter on
each endpoint can be set to true
. This runs the chaos process in a Sidekiq worker.
Memory leaks
To simulate a memory leak in your application, use the /-/chaos/leakmem
endpoint.
The memory is not retained after the request finishes. After the request has completed, the Ruby garbage collector attempts to recover the memory.
GET /-/chaos/leakmem
GET /-/chaos/leakmem?memory_mb=1024
GET /-/chaos/leakmem?memory_mb=1024&duration_s=50
GET /-/chaos/leakmem?memory_mb=1024&duration_s=50&async=true
Attribute | Type | Required | Description |
---|---|---|---|
memory_mb |
integer | no | How much memory, in MB, should be leaked. Defaults to 100MB. |
duration_s |
integer | no | Minimum duration_s, in seconds, that the memory should be retained. Defaults to 30s. |
async |
boolean | no | Set to true to leak memory in a Sidekiq background worker process |
curl "http://localhost:3000/-/chaos/leakmem?memory_mb=1024&duration_s=10" \
--header 'X-Chaos-Secret: secret'
curl "http://localhost:3000/-/chaos/leakmem?memory_mb=1024&duration_s=10&token=secret"
CPU spin
This endpoint attempts to fully utilise a single core, at 100%, for the given period.
Depending on your rack server setup, your request may timeout after a predetermined period (normally 60 seconds). If you're using Unicorn, this is done by killing the worker process.
GET /-/chaos/cpu_spin
GET /-/chaos/cpu_spin?duration_s=50
GET /-/chaos/cpu_spin?duration_s=50&async=true
Attribute | Type | Required | Description |
---|---|---|---|
duration_s |
integer | no | Duration, in seconds, that the core is used. Defaults to 30s |
async |
boolean | no | Set to true to consume CPU in a Sidekiq background worker process |
curl "http://localhost:3000/-/chaos/cpu_spin?duration_s=60" \
--header 'X-Chaos-Secret: secret'
curl "http://localhost:3000/-/chaos/cpu_spin?duration_s=60&token=secret"
DB spin
This endpoint attempts to fully utilise a single core, and interleave it with DB request, for the given period. This endpoint can be used to model yielding execution to another threads when running concurrently.
Depending on your rack server setup, your request may timeout after a predetermined period (normally 60 seconds). If you're using Unicorn, this is done by killing the worker process.
GET /-/chaos/db_spin
GET /-/chaos/db_spin?duration_s=50
GET /-/chaos/db_spin?duration_s=50&async=true
Attribute | Type | Required | Description |
---|---|---|---|
interval_s |
float | no | Interval, in seconds, for every DB request. Defaults to 1s |
duration_s |
integer | no | Duration, in seconds, that the core is used. Defaults to 30s |
async |
boolean | no | Set to true to perform the operation in a Sidekiq background worker process |
curl "http://localhost:3000/-/chaos/db_spin?interval_s=1&duration_s=60" \
--header 'X-Chaos-Secret: secret'
curl "http://localhost:3000/-/chaos/db_spin?interval_s=1&duration_s=60&token=secret"
Sleep
This endpoint is similar to the CPU Spin endpoint but simulates off-processor activity, such as network calls to backend services. It sleeps for a given duration_s
.
As with the CPU Spin endpoint, this may lead to your request timing out if duration_s
exceeds the configured limit.
GET /-/chaos/sleep
GET /-/chaos/sleep?duration_s=50
GET /-/chaos/sleep?duration_s=50&async=true
Attribute | Type | Required | Description |
---|---|---|---|
duration_s |
integer | no | Duration, in seconds, that the request sleeps for. Defaults to 30s |
async |
boolean | no | Set to true to sleep in a Sidekiq background worker process |
curl "http://localhost:3000/-/chaos/sleep?duration_s=60" \
--header 'X-Chaos-Secret: secret'
curl "http://localhost:3000/-/chaos/sleep?duration_s=60&token=secret"
Kill
This endpoint simulates the unexpected death of a worker process using the KILL
signal.
Because this endpoint uses the KILL
signal, the process isn't given an
opportunity to clean up or shut down.
GET /-/chaos/kill
GET /-/chaos/kill?async=true
Attribute | Type | Required | Description |
---|---|---|---|
async |
boolean | no | Set to true to signal a Sidekiq background worker process |
curl "http://localhost:3000/-/chaos/kill" --header 'X-Chaos-Secret: secret'
curl "http://localhost:3000/-/chaos/kill?token=secret"
Quit
This endpoint simulates the unexpected death of a worker process using the QUIT
signal.
Unlike KILL
, the QUIT
signal will also attempt to write a core dump.
See core(5) for more information.
GET /-/chaos/quit
GET /-/chaos/quit?async=true
Attribute | Type | Required | Description |
---|---|---|---|
async |
boolean | no | Set to true to signal a Sidekiq background worker process |
curl "http://localhost:3000/-/chaos/quit" --header 'X-Chaos-Secret: secret'
curl "http://localhost:3000/-/chaos/quit?token=secret"
Run garbage collector
This endpoint triggers a GC run on the worker handling the request and returns its worker ID plus GC stats as JSON. This is mostly useful when running Puma in standalone mode, since otherwise the worker handling the request will not be known upfront.
Endpoint:
POST /-/chaos/gc
Example request:
curl --request POST "http://localhost:3000/-/chaos/gc" \
--header 'X-Chaos-Secret: secret'
curl --request POST "http://localhost:3000/-/chaos/gc?token=secret"
Example response:
{
"worker_id": "puma_1",
"gc_stat": {
"count": 94,
"heap_allocated_pages": 9077,
"heap_sorted_length": 9077,
"heap_allocatable_pages": 0,
"heap_available_slots": 3699720,
"heap_live_slots": 2827510,
"heap_free_slots": 872210,
"heap_final_slots": 0,
"heap_marked_slots": 2827509,
"heap_eden_pages": 9077,
"heap_tomb_pages": 0,
"total_allocated_pages": 9077,
"total_freed_pages": 0,
"total_allocated_objects": 14229357,
"total_freed_objects": 11401847,
"malloc_increase_bytes": 8192,
"malloc_increase_bytes_limit": 30949538,
"minor_gc_count": 71,
"major_gc_count": 23,
"compact_count": 0,
"remembered_wb_unprotected_objects": 41685,
"remembered_wb_unprotected_objects_limit": 83370,
"old_objects": 2617806,
"old_objects_limit": 5235612,
"oldmalloc_increase_bytes": 8192,
"oldmalloc_increase_bytes_limit": 122713697
}
}