We started with the most battle-tested and native option to Postgres, which is PgBouncer and tried tuning it the right way. Also now that long due kinks like support for prepared statements are solved, it’s been working really well. There are many customers scaling well with 10K+ Postgres connections. We will consider other options like odyssey, pgdog in the future!
Side note: I’m not a big fan of having 10K+ connections on Postgres, 100s are more than enough to scale Postgres well. But that’s a story for another day. ;)
Cool, another whataboutism. USA caused a lot of damage, but USA is a grey country - they do a lot of bad things but they also do a lot of good things (science & technology).
russia is one-dimensional and that dimension is violence. They don't invent things, they don't innovate, they don't export anything (apart from gas/minerals) and they only have imperialistic ambitions.
> The cancel lands on a process that has never heard of the query, and nothing happens.
> Peering fixes this. The processes are aware of one another, so a cancel that lands on the wrong process is forwarded to the one that actually owns the session.
I understand "peering" as a concept here but have never tried this with PostgreSQL before. May I ask:
A) Does PostgreSQL have a mode/setting for peering that makes this easy? I'm imagining a mechanism that either goes round robin (re-sending the cancel to peers until it doesn't return an error of some kind) or some metadata in the cancel request that enables the wrong-destination process to somehow identify the proper process.
B) And by what mechanism? If all the PostgreSQL processes are listening to clients via so_reuseport, I guess there must be some other IPC method used for the peering chatter.
AI clearly wrote TFA. The cancellation thing is apparently a PgBouncer feature - the peering is between bouncer processes, not server processes. It sounds like it should be easy enough to make the bouncer process ID part of the cancel key.
Interesting. We run pgbouncer via kubernetes so it was straightforward to make multiple pgbouncer processes on one machine. Also straightforward to get them running on multiple machines, which helps because we run on Azure and they like to cause rolling outages across our fleet via VM maintenance...
Ack, makes sense. I’m very curious on how this affects throughput due to a potential extra network hop from pgbouncer to Postgres. Expecting it to have a minor difference, but still curious.
Glad you like it. It was built to fix some of PgBouncers shortcomings that we ran into at Instacart many years ago, and to have a stronger foundation for scaling Postgres horizontally (that's sharding)!
First time I've heard of so_reuseport, which is interesting. The important parts of the setup seem to be that + peering; is peering built-in to PgBouncer and simple to set up?
However, if pooling isn’t used, there’s always an overhead (tens of milliseconds or more) when creating a new connection because Postgres needs to fork a process. And yes, applications can be written without pooling, which isn’t ideal, but happens quite a lot.
Application frameworks have also changed. Serverless architectures can generate tens of thousands of connections, which is where Postgres starts to run into issues. I’m personally not a big fan of using more than a few hundred connections, but it is very realistic in this era.
This was more for fun than real use, but I greatly enjoyed hacking something similar into rqbit bittorrent client.
I wanted to run an instance of 'rqbit download' per torrent via so_reuseport. When a peer tries to connect, it gets sent to a random instance. So I built a whole rendezvous system, where instances find each other & either proxy data to each other or fd pass the socket to each other directly to get the peer socket to the instance that needs it. It uses postcard rpc to chat between instances.
Clickhouse's so_reuseport rendezvous needs are obviously for a very different, but fun to see some so_reuseport coordination like this (for a much more practical use)!
It'd be really neat to have some kind of general peering protocol that different apps could use. This whole exercise was gratuitous as heck for my application, I don't even really intend to use this, but it was a fun path to walk down. So I don't really know what the broader protocol would really be for, what we would use it for. But it seems like such a cool idea! A shared Turso database would probably be a bit more practical than the rpc system, honestly. Ha.
Just use https://github.com/yandex/odyssey :) It's a scalable PgBouncer.
We started with the most battle-tested and native option to Postgres, which is PgBouncer and tried tuning it the right way. Also now that long due kinks like support for prepared statements are solved, it’s been working really well. There are many customers scaling well with 10K+ Postgres connections. We will consider other options like odyssey, pgdog in the future!
Side note: I’m not a big fan of having 10K+ connections on Postgres, 100s are more than enough to scale Postgres well. But that’s a story for another day. ;)
Fun (semi-related) fact, ClickHouse was originally developed by Yandex :)
yandex is a russian company, so no, don't use it
Do you boycott American companies too? The USA caused much more death and destruction than Russia ever did.
Cool, another whataboutism. USA caused a lot of damage, but USA is a grey country - they do a lot of bad things but they also do a lot of good things (science & technology).
russia is one-dimensional and that dimension is violence. They don't invent things, they don't innovate, they don't export anything (apart from gas/minerals) and they only have imperialistic ambitions.
uhhhh
nginx is also Russian, so don't use... a third of the web.
nginx was built by one person. I'm talking about a russian corporation that has (inescapable) ties to kremlin
I have bad news for you about where ClickHouse comes from then.
thank you for the advice mr mcarthy, i'll happily use deepseek and GLM to vibe my next project.
> The cancel lands on a process that has never heard of the query, and nothing happens.
> Peering fixes this. The processes are aware of one another, so a cancel that lands on the wrong process is forwarded to the one that actually owns the session.
I understand "peering" as a concept here but have never tried this with PostgreSQL before. May I ask:
A) Does PostgreSQL have a mode/setting for peering that makes this easy? I'm imagining a mechanism that either goes round robin (re-sending the cancel to peers until it doesn't return an error of some kind) or some metadata in the cancel request that enables the wrong-destination process to somehow identify the proper process.
B) And by what mechanism? If all the PostgreSQL processes are listening to clients via so_reuseport, I guess there must be some other IPC method used for the peering chatter.
AI clearly wrote TFA. The cancellation thing is apparently a PgBouncer feature - the peering is between bouncer processes, not server processes. It sounds like it should be easy enough to make the bouncer process ID part of the cancel key.
Interesting. We run pgbouncer via kubernetes so it was straightforward to make multiple pgbouncer processes on one machine. Also straightforward to get them running on multiple machines, which helps because we run on Azure and they like to cause rolling outages across our fleet via VM maintenance...
Ack, makes sense. I’m very curious on how this affects throughput due to a potential extra network hop from pgbouncer to Postgres. Expecting it to have a minor difference, but still curious.
Cross zone latency has noticeable effect: https://news.ycombinator.com/item?id=45512351
I've been using pgdog (https://github.com/pgdogdev/pgdog) and it has worked really well for my needs!
Glad you like it. It was built to fix some of PgBouncers shortcomings that we ran into at Instacart many years ago, and to have a stronger foundation for scaling Postgres horizontally (that's sharding)!
Ack, we will consider offering pgdog in the future! More context on why we chose pgbouncer https://news.ycombinator.com/item?id=48873867
It's been a pong pong back and forth between PG reverse proxy submissions lately! "Why we built [Pgdog]": https://news.ycombinator.com/item?id=48819308
First time I've heard of so_reuseport, which is interesting. The important parts of the setup seem to be that + peering; is peering built-in to PgBouncer and simple to set up?
Yes, https://www.pgbouncer.org/config.html#section-peers
I'm 46 now. I remember being shocked at Postgres's heavy connection model when I was 23.
I gather things haven't improved since?
It improved quite a lot! It scales pretty well to thousands of connections: https://techcommunity.microsoft.com/blog/adforpostgresql/imp....
However, if pooling isn’t used, there’s always an overhead (tens of milliseconds or more) when creating a new connection because Postgres needs to fork a process. And yes, applications can be written without pooling, which isn’t ideal, but happens quite a lot.
Application frameworks have also changed. Serverless architectures can generate tens of thousands of connections, which is where Postgres starts to run into issues. I’m personally not a big fan of using more than a few hundred connections, but it is very realistic in this era.
Was there a disadvantage to using HAProxy + multiple PGBouncer instances?
SO_REUSEPORT[1] pretty much does all we want in kernelspace vs unnecessary userspace hop inbetween. These all run on same VM
1: https://lwn.net/Articles/542629
This was more for fun than real use, but I greatly enjoyed hacking something similar into rqbit bittorrent client. I wanted to run an instance of 'rqbit download' per torrent via so_reuseport. When a peer tries to connect, it gets sent to a random instance. So I built a whole rendezvous system, where instances find each other & either proxy data to each other or fd pass the socket to each other directly to get the peer socket to the instance that needs it. It uses postcard rpc to chat between instances.
Clickhouse's so_reuseport rendezvous needs are obviously for a very different, but fun to see some so_reuseport coordination like this (for a much more practical use)!
It'd be really neat to have some kind of general peering protocol that different apps could use. This whole exercise was gratuitous as heck for my application, I don't even really intend to use this, but it was a fun path to walk down. So I don't really know what the broader protocol would really be for, what we would use it for. But it seems like such a cool idea! A shared Turso database would probably be a bit more practical than the rpc system, honestly. Ha.
https://github.com/rektide/rqbit/tree/peering
Article should show the config:
[pgbouncer] listen_addr = 0.0.0.0 listen_port = 6432 so_reuseport = 1 peer_id = 1 unix_socket_dir = /tmp/pgbouncer1
[peers] 1 = host=/tmp/pgbouncer1 2 = host=/tmp/pgbouncer2 3 = host=/tmp/pgbouncer3 4 = host=/tmp/pgbouncer4