Reproducible benchmark for PostgREST by using Nix and NixOps.
NixOps provisions AWS EC2 instances on a dedicated VPC and deploys the different components for load testing.
The default setup includes:
- A
m5a.xlarge
instance which uses k6 for load testing. - A
t3a.nano
instance with PostgreSQL. - A
t3a.nano
instance with PostgREST + Nginx.
This setup, including the size of the EC2 instances, can be modified with environment variables. As the PostgreSQL server instance size increases, its settings are modified according to PGTune recommendations.
-
AWS account with an
~/.aws/credentials
file in place (can be created withaws-cli
). The "default" profile is picked up by default but you can change it by doing:export PGRSTBENCH_AWS_PROFILE="another_profile"
-
Nix.
Run nix-shell
. This will provide an environment where all the dependencies are available.
$ nix-shell
>
Deploy with:
$ postgrest-bench-deploy
pgrstBenchVpc.....> creating vpc under region us-east-2
..
pg................> activation finished successfully
pgrstbench> deployment finished successfully
# this command will take a couple minutes, it will deploy the client and server AWS machines VPC stuff
Run a k6
test on the client instance and get the output:
$ postgrest-bench-k6 20 k6/GETSingle.js
Destroy all the setup and the AWS instances:
$ postgrest-bench-destroy
To connect to the PostgreSQL instance:
$ postgrest-bench-ssh pg
# Check the installed services
$ systemctl list-units
## connect to postgres
$ psql -U postgres
$ \d
The postgresql server comes loaded with the chinook database.
To connect to the PostgREST instance:
$ postgrest-bench-ssh pgrst
# Check the installed services
$ systemctl list-units
# Do a request
$ curl localhost:80/artist
You can also get info (like the IPs) of the instances with:
$ postgrest-bench-info
K6 runs on the client instance, but you can get the output of the load test on your machine:
## k6 will run with 10 VUs on the AWS client instance and load test the t3anano instance with the local k6/GETSingle.js script
$ pgrbench-k6 10 k6/GETSingle.js
## You will see the k6 logo and runs here
There are different scripts on k6/
which test different PostgREST requests.
pgbench also runs on the client instance, you can get its output with:
$ postgrest-bench-pgbench pgbench/GETSingle.sql
The GETSingle.sql
runs an equivalent SQL statement to what PostgREST generates for GETSingle.js
. The motivation for this comparison is to see how much PostgREST performance differs from direct SQL connections.
There are scripts that help with varying the environment while load testing. You can use these to get a report once the command finishes running:
Run pgbench with a different qty of clients:
$ postgrest-bench-pgbench-vary-clients pgbench/GETSingle.sql
Run k6 with a different qty of VUs:
$ postgrest-bench-k6-vary-vus k6/GETSingle.js
Run pgbench with varied clients and with varied pg instances (this will involve reprovisioning/redeploying the pg instance, it will take a while):
$ postgrest-bench-vary-pg postgrest-bench-pgbench-vary-clients pgbench/GETSingle.sql > PGBENCH_GET_SINGLE.txt
Run k6 with varied vus and with varied pg instances and pgrst instances (this will involve reprovisioning/redeploying the pg and pgrst instance, it will take even longer):
$ postgrest-bench-vary-pg-pgrst postgrest-bench-k6-vary-vus k6/GETSingle.js > K6_GET_SINGLE.txt
To load test with nginx included do:
export PGRBENCH_WITH_NGINX="true"
postgrest-bench-deploy
To only have PostgREST listening directly on port 80:
export PGRBENCH_WITH_NGINX="false"
postgrest-bench-deploy
To load test connecting pgrest to pg with unix socket, and pgrest to nginx with unix socket.
export PGRBENCH_WITH_UNIX_SOCKET="true"
postgrest-bench-deploy
To use tcp instead, you can do:
export PGRBENCH_WITH_UNIX_SOCKET="false"
postgrest-bench-deploy
To load test with a pg on a different ec2 instance.
export PGRBENCH_SEPARATE_PG="true"
postgrest-bench-deploy
To use the same instance for both PostgreSQL and PostgREST.
export PGRBENCH_SEPARATE_PG="false"
postgrest-bench-deploy
Some experiments indicate that when load testing on big instances (like m5a.8xlarge
and above), having two postgrest instances instead of one increase throughput. You can try this with:
export PGRSTBENCH_PGRST_NGNIX_LBS="true"
postgrest-bench-deploy
To change pg and PostgREST EC2 instance types(both t3a.nano
by default):
export PGRBENCH_PG_INSTANCE_TYPE="t3a.xlarge"
export PGRBENCH_PGRST_INSTANCE_TYPE="t3a.xlarge"
export PGRBENCH_PGRST_INSTANCE_TYPE="t3a.xlarge"
postgrest-bench-deploy
- Uses an outdated version of NixOps, 1.7. Newer versions have changed considerably.
- Don't try changing to ARM-based instances, these don't work with the above NixOps version.
- The instances tested for this benchmark are the
t3a
series and them5a
series.
- The instances tested for this benchmark are the
- majkinetor/postgrest-test: PostgREST benchmark on Windows.