Meow. PgBouncer rewritten in Rust, with sharding, load balancing and failover support.
Alpha: don't use in production just yet.
- Install Rust (latest stable will work great).
cargo run --release(to get better benchmarks).- Change the config in
pgcat.tomlto fit your setup (optional given next step). - Install Postgres and run
psql -f tests/sharding/query_routing_setup.sql
You can just PgBench to test your changes:
pgbench -i -h 127.0.0.1 -p 6432 && \
pgbench -t 1000 -p 6432 -h 127.0.0.1 --protocol simple && \
pgbench -t 1000 -p 6432 -h 127.0.0.1 --protocol extended
See sharding README for sharding logic testing.
- Session mode.
- Transaction mode.
COPYprotocol support.- Query cancellation.
- Round-robin load balancing of replicas.
- Banlist & failover.
- Sharding!
- Explicit query routing to primary or replicas.
Each client owns its own server for the duration of the session. Commands like SET are allowed.
This is identical to PgBouncer session mode.
The connection is attached to the server for the duration of the transaction. SET will pollute the connection,
but SET LOCAL works great. Identical to PgBouncer transaction mode.
That one isn't particularly special, but good to mention that you can COPY data in and from the server
using this pooler.
Okay, this is just basic stuff, but we support cancelling queries. If you know the Postgres protocol, this might be relevant given than this is a transactional pooler but if you're new to Pg, don't worry about it, it works.
This is the novel part. PgBouncer doesn't support it and suggests we use DNS or a TCP proxy instead. We prefer to have everything as part of one package; arguably, it's easier to understand and optimize. This pooler will round-robin between multiple replicas keeping load reasonably even. If the primary is in the pool as well, it'll be treated as a replica for read-only queries.
This is where it gets even more interesting. If we fail to connect to one of the replicas or it fails a health check, we add it to a ban list. No more new transactions will be served by that replica for, in our case, 60 seconds. This gives it the opportunity to recover while clients are happily served by the remaining replicas.
This decreases error rates substantially! Worth noting here that on busy systems, if the replicas are running too hot, failing over could bring even more load and tip over the remaining healthy-ish replicas. In this case, a decision should be made: either lose 1/x of your traffic or risk losing it all eventually. Ideally you overprovision your system, so you don't necessarily need to make this choice :-).
We're implemeting Postgres' PARTITION BY HASH sharding function for BIGINT fields. This works well for tables that use BIGSERIAL primary key which I think is common enough these days. We can also add many more functions here, but this is a good start. See src/sharding.rs and tests/sharding/partition_hash_test_setup.sql for more details on the implementation.
The biggest advantage of using this sharding function is that anyone can shard the dataset using Postgres partitions while also access it for both reads and writes using this pooler. No custom obscure sharding function is needed and database sharding can be done entirely in Postgres.
To select the shard we want to talk to, we introduced special syntax:
SET SHARDING KEY TO '1234';This sharding key will be hashed and the pooler will select a shard to use for the next transaction. If the pooler is in session mode, this sharding key has to be set as the first query on startup & cannot be changed until the client re-connects.
If you want to have the primary and replicas in the same pooler, you'd probably want to
route queries explicitely to the primary or replicas, depending if they are reads or writes (e.g SELECTs or INSERT/UPDATE, etc). To help with this, we introduce some more custom syntax:
SET SERVER ROLE TO 'primary';
SET SERVER ROLE TO 'replica';After executing this, the next transaction will be routed to the primary or replica respectively. By default, all queries will be load-balanced between all servers, so if the client wants to write or talk to the primary, they have to explicitely select it using the syntax above.
- Authentication, ehem, this proxy is letting anyone in at the moment.
You can setup PgBench locally through PgCat:
pgbench -h 127.0.0.1 -p 6432 -i
Coincidenly, this uses COPY so you can test if that works. Additionally, we'll be running the following PgBench configurations:
- 16 clients, 2 threads
- 32 clients, 2 threads
- 64 clients, 2 threads
- 128 clients, 2 threads
All queries will be SELECT only (-S) just so disks don't get in the way, since the dataset will be effectively all in RAM.
My setup:
- 8 cores, 16 hyperthreaded (AMD Ryzen 5800X)
- 32GB RAM (doesn't matter for this benchmark, except to prove that Postgres will fit the whole dataset into RAM)
[databases]
shard0 = host=localhost port=5432 user=sharding_user password=sharding_user
[pgbouncer]
pool_mode = transaction
max_client_conn = 1000Everything else stays default.
$ pgbench -t 1000 -c 16 -j 2 -p 6432 -h 127.0.0.1 -S --protocol extended shard0
starting vacuum...end.
transaction type: <builtin: select only>
scaling factor: 1
query mode: extended
number of clients: 16
number of threads: 2
number of transactions per client: 1000
number of transactions actually processed: 16000/16000
latency average = 0.155 ms
tps = 103417.377469 (including connections establishing)
tps = 103510.639935 (excluding connections establishing)
$ pgbench -t 1000 -c 32 -j 2 -p 6432 -h 127.0.0.1 -S --protocol extended shard0
starting vacuum...end.
transaction type: <builtin: select only>
scaling factor: 1
query mode: extended
number of clients: 32
number of threads: 2
number of transactions per client: 1000
number of transactions actually processed: 32000/32000
latency average = 0.290 ms
tps = 110325.939785 (including connections establishing)
tps = 110386.513435 (excluding connections establishing)
$ pgbench -t 1000 -c 64 -j 2 -p 6432 -h 127.0.0.1 -S --protocol extended shard0
starting vacuum...end.
transaction type: <builtin: select only>
scaling factor: 1
query mode: extended
number of clients: 64
number of threads: 2
number of transactions per client: 1000
number of transactions actually processed: 64000/64000
latency average = 0.692 ms
tps = 92470.427412 (including connections establishing)
tps = 92618.389350 (excluding connections establishing)
$ pgbench -t 1000 -c 128 -j 2 -p 6432 -h 127.0.0.1 -S --protocol extended shard0
starting vacuum...end.
transaction type: <builtin: select only>
scaling factor: 1
query mode: extended
number of clients: 128
number of threads: 2
number of transactions per client: 1000
number of transactions actually processed: 128000/128000
latency average = 1.406 ms
tps = 91013.429985 (including connections establishing)
tps = 91067.583928 (excluding connections establishing)
The only thing that matters here is the number of workers in the Tokio pool. Make sure to set it to < than the number of your CPU cores. Also account for hyper-threading, so if you have that, take the number you got above and divide it by two, that way only "real" cores serving requests.
My setup is 16 threads, 8 cores (htop shows as 16 CPUs), so I set the max_workers in Tokio to 4. Too many, and it starts conflicting with PgBench
which is also running on the same system.
$ pgbench -t 1000 -c 16 -j 2 -p 6432 -h 127.0.0.1 -S --protocol extended
starting vacuum...end.
transaction type: <builtin: select only>
scaling factor: 1
query mode: extended
number of clients: 16
number of threads: 2
number of transactions per client: 1000
number of transactions actually processed: 16000/16000
latency average = 0.164 ms
tps = 97705.088232 (including connections establishing)
tps = 97872.216045 (excluding connections establishing)
$ pgbench -t 1000 -c 32 -j 2 -p 6432 -h 127.0.0.1 -S --protocol extended
starting vacuum...end.
transaction type: <builtin: select only>
scaling factor: 1
query mode: extended
number of clients: 32
number of threads: 2
number of transactions per client: 1000
number of transactions actually processed: 32000/32000
latency average = 0.288 ms
tps = 111300.488119 (including connections establishing)
tps = 111413.107800 (excluding connections establishing)
$ pgbench -t 1000 -c 64 -j 2 -p 6432 -h 127.0.0.1 -S --protocol extended
starting vacuum...end.
transaction type: <builtin: select only>
scaling factor: 1
query mode: extended
number of clients: 64
number of threads: 2
number of transactions per client: 1000
number of transactions actually processed: 64000/64000
latency average = 0.556 ms
tps = 115190.496139 (including connections establishing)
tps = 115247.521295 (excluding connections establishing)
$ pgbench -t 1000 -c 128 -j 2 -p 6432 -h 127.0.0.1 -S --protocol extended
starting vacuum...end.
transaction type: <builtin: select only>
scaling factor: 1
query mode: extended
number of clients: 128
number of threads: 2
number of transactions per client: 1000
number of transactions actually processed: 128000/128000
latency average = 1.135 ms
tps = 112770.562239 (including connections establishing)
tps = 112796.502381 (excluding connections establishing)
Always good to have a base line.
$ pgbench -t 1000 -c 16 -j 2 -p 5432 -h 127.0.0.1 -S --protocol extended shard0
Password:
starting vacuum...end.
transaction type: <builtin: select only>
scaling factor: 1
query mode: extended
number of clients: 16
number of threads: 2
number of transactions per client: 1000
number of transactions actually processed: 16000/16000
latency average = 0.115 ms
tps = 139443.955722 (including connections establishing)
tps = 142314.859075 (excluding connections establishing)
$ pgbench -t 1000 -c 32 -j 2 -p 5432 -h 127.0.0.1 -S --protocol extended shard0
Password:
starting vacuum...end.
transaction type: <builtin: select only>
scaling factor: 1
query mode: extended
number of clients: 32
number of threads: 2
number of transactions per client: 1000
number of transactions actually processed: 32000/32000
latency average = 0.212 ms
tps = 150644.840891 (including connections establishing)
tps = 152218.499430 (excluding connections establishing)
$ pgbench -t 1000 -c 64 -j 2 -p 5432 -h 127.0.0.1 -S --protocol extended shard0
Password:
starting vacuum...end.
transaction type: <builtin: select only>
scaling factor: 1
query mode: extended
number of clients: 64
number of threads: 2
number of transactions per client: 1000
number of transactions actually processed: 64000/64000
latency average = 0.420 ms
tps = 152517.663404 (including connections establishing)
tps = 153319.188482 (excluding connections establishing)
$ pgbench -t 1000 -c 128 -j 2 -p 5432 -h 127.0.0.1 -S --protocol extended shard0
Password:
starting vacuum...end.
transaction type: <builtin: select only>
scaling factor: 1
query mode: extended
number of clients: 128
number of threads: 2
number of transactions per client: 1000
number of transactions actually processed: 128000/128000
latency average = 0.854 ms
tps = 149818.594087 (including connections establishing)
tps = 150200.603049 (excluding connections establishing)
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent