/pg-gateway

Postgres wire protocol for the server-side

Primary LanguageTypeScriptMIT LicenseMIT

pg-gateway

TypeScript library that implements the Postgres wire protocol from the server-side. It provides APIs you can hook into to handle authentication requests, queries, and other client messages yourself.

Why?

This acts as a layer in front of your Postgres database (or any other database). You could use it for:

  • Serving PGlite over TCP
  • Serving a non-Postgres database via the Postgres wire protocol
  • Adding a reverse proxy in front of your databases
  • Creating a custom connection pooler

Features

  • Authentication: Supports multiple auth modes - currently cleartextPassword, md5Password, and certificate (more planned)
  • TLS Encryption: Handles standard TLS (SSL) upgrades with SNI support (useful for reverse proxying)
  • Modular: You control the server while the library manages the protocol
  • Hooks: Hook into various points in the protocol's lifecycle (auth, query, etc)
  • Escape hatch: Access the raw protocol messages at any point in the lifecycle
  • Examples: Various examples on how you might use this library

Usage

This library is in active development, so APIs are still WIP. It is pre-1.0 so expect some breaking changes to the API.

This library is designed to give you as much control as possible while still managing the protocol lifecycle.

Start by creating your own TCP server (ie. via node:net), then pass the socket into a PostgresConnection:

import { createServer } from 'node:net';
import { PostgresConnection } from 'pg-gateway';

// Create a TCP server
const server = createServer((socket) => {
  // `PostgresConnection` will manage the protocol lifecycle
  const connection = new PostgresConnection(socket);
});

// Listen on the desired port
server.listen(5432, () => {
  console.log('Server listening on port 5432');
});

PostgresConnection exposes a number of options and hooks as its second argument:

serverVersion

Specifies the version of the server to return back to the client. Can include any arbitrary string.

const connection = new PostgresConnection(socket, {
  serverVersion: '16.3 (MyCustomPG)',
});

authMode

Specifies which auth mode you want to use with clients. Current modes supported are:

  • none: No password is requested. Client will be immediately authenticate after startup.
  • cleartextPassword: Password is sent in plain text by the client. Least secure option.
  • md5Password: Password is hashed using Postgres' nested MD5 algorithm before it is sent to the server.
  • certificate: A client TLS certificate is requested. Its common name (CN) must match the user field to authenticate. It must also be signed by a certificate authority (CA) known to the server (see TLS). Requires a TLS connection.
const connection = new PostgresConnection(socket, {
  authMode: `md5Password`,
});

tls

Like the real Postgres server, TLS connections are established as an upgrade mechanism after the initial handshake (SSLRequest message from the client).

The tls option is an object that contains the following:

  • key: A Buffer containing the TLS private key in PEM format
  • cert: A Buffer containing the TLS cert chain in PEM format. Includes the TLS cert followed by any intermediate certs (not including root CA).
  • ca: A Buffer containing the certificate authority (CA) in PEM format. Optional - if omitted, the runtime's built-in CAs will be used.
  • passphrase A string containing the passphrase for key. Optional - only needed if key is encrypted.

When this option is passed, the server will require a TLS connection with the client. If the client doesn't send an SSLRequest message, the server will close the connection with an error.

If this option is not passed, the server will respond to SSLRequest messages with a message stating that SSL is not supported. At that point, clients can decide whether or not they want to continue the connection over an unencrypted channel.

const tls: TlsOptions = {
  key: readFileSync('server-key.pem'),
  cert: readFileSync('server-cert.pem'),
  ca: readFileSync('ca-cert.pem'),
};

const connection = new PostgresConnection(socket, {
  tls,
});

onTlsUpgrade()

This hook is called after the TLS upgrade has completed. It passes a state argument which holds connection information gathered so far like tlsInfo. The callback can be either synchronous or asynchronous.

This will be called before the startup message is received from the frontend (if TLS is being used) so is a good place to establish proxy connections if desired. Note that a detach() method is also available if you wish to detach from the PostgresConnection after the proxy has been established.

const tls: TlsOptions = {
  key: readFileSync('server-key.pem'),
  cert: readFileSync('server-cert.pem'),
  ca: readFileSync('ca-cert.pem'),
};

const connection = new PostgresConnection(socket, {
  tls,
  async onTlsUpgrade({ tlsInfo }) {
    console.log({ tlsInfo });
  },
});

onStartup()

This hook is called after the initial startup message has been received from the frontend. It passes a state argument which holds connection information gathered so far like clientInfo. The callback can be either synchronous or asynchronous.

This is called after the connection is upgraded to TLS (if TLS is being used) but before authentication messages are sent to the frontend.

The callback should return true to indicate that it has responded to the startup message and no further processing should occur. This can be useful for situations where you want to receive the startup client information (such as user), but then proxy all future messages (including auth messages) to another DB. Return false to continue built-in processing.

Warning: By managing the post-startup response yourself (returning true), you bypass further processing by the PostgresConnection which means some state may not be collected and hooks won't be called.

const connection = new PostgresConnection(socket, {
  async onStartup({ clientInfo }) {
    console.log({ clientInfo });
  },
});

onAuthenticated()

This hook is called after a successful authentication has completed. It passes a state argument which holds connection information gathered so far. The callback can be either synchronous or asynchronous.

const connection = new PostgresConnection(socket, {
  async onAuthenticated(state) {
    console.log(state);
  },
});

validateCredentials()

This hook allows you to authenticate credentials based on the auth mode. Returning true indicates that the credentials are valid and false indicates that they are invalid. If the credentials are marked invalid, the server will close the connection with an error. The callback can be either synchronous or asynchronous.

  1. The first argument contains a credentials object passed to the function will contain different properties based on the auth mode:

    • cleartextPassword:

      • authMode: 'cleartextPassword';
      • user: string;
      • password: string;

    • md5Password:

      • authMode: 'md5Password';
      • user: string;
      • hash: string;
      • salt: Uint8Array;

  2. The second argument contains a state object which holds connection information gathered so far and can be used to understand where the protocol is at in its lifecycle.

You can use TypeScript's discriminated unions to narrow the type of credentials before accessing its properties:

import { PostgresConnection, hashMd5Password } from 'pg-gateway';

// ...

const connection = new PostgresConnection(socket, {
  async validateCredentials(credentials) {
    if (credentials.authMode === 'md5Password') {
      const { hash, salt } = credentials;
      const expectedHash = await hashMd5Password('postgres', 'postgres', salt);
      return hash === expectedHash;
    }
    return false;
  },
});

onMessage()

This hook gives you access to raw messages at any point in the protocol lifecycle.

  1. The first argument contains the raw Buffer data in the message
  2. The second argument contains a state object which holds connection information gathered so far and can be used to understand where the protocol is at in its lifecycle.

The callback should return true to indicate that you have handled the message response yourself and that no further processing should be done. Returning false will result in further processing by the PostgresConnection. The callback can be either synchronous or asynchronous.

Warning: By managing the message yourself (returning true), you bypass further processing by the PostgresConnection which means some state may not be collected and hooks won't be called depending on where the protocol is at in its lifecycle.

const connection = new PostgresConnection(socket, {
  async onMessage(data, { hasStarted, isAuthenticated }) {
    // Handle raw messages yourself

    return false;
  },
});

See PGlite for an example on how you might use this.

State

Over the course of the protocol lifecycle, pg-gateway will hold a state object that consists of various connection information gathered. Below are the properties available in state:

  • hasStarted: boolean indicating whether or not a startup message has been received by the client

  • isAuthenticated: boolean indicating whether or not a successful authentication handshake has completed with the client

  • clientInfo: object containing client information sent during startup.

    • majorVersion: number;
    • minorVersion: number;
    • parameters:
      • user: the user to connect to the database with
      • Any other arbitrary key-value pair (often database and run-time parameters). See StartupMessage for more info.

    Note that clientInfo will be undefined until a startup message is received.

  • tlsInfo: object containing TLS connection information (if TLS is being used).

    • sniServerName: string containing the SNI server name sent by the client. This can be undefined if the SNI extension was not used by the client.

    Note that tlsInfo will be undefined until a TLS upgrade has completed.

State is available directly on the PostgresConnection instance:

const connection = new PostgresConnection(socket);
console.log(connection.state);

It is also passed as an argument to most hooks for convenience.

detach()

A detach() method exists on the PostgresConnection to allow you to completely detach the socket from the PostgresConnection and handle all future data processing yourself. This is useful when reverse proxying to prevent the PostgresConnection from continuing to process each message after the proxy connection is established.

Calling detach() will return the current Socket which may be different than the original socket if a TLS upgrade occurred (ie. a TLSSocket). The PostgresConnection will remove all event listeners from the socket and no further processing will take place.

const connection = new PostgresConnection(socket);
const socket = connection.detach();

Examples

PGlite

PGlite is a WASM build of Postgres that can run in-browser or server-side. Under the hood, PGlite uses Postgres' single-user mode which skips standard startup/auth messages in the protocol and operates outside of a TCP connection.

With pg-gateway, we can serve PGlite over TCP by handling the startup/auth ourselves, then forward future messages to PGlite:

import { PGlite } from '@electric-sql/pglite';
import net from 'node:net';
import { PostgresConnection } from 'pg-gateway';

const db = new PGlite();

const server = net.createServer((socket) => {
  const connection = new PostgresConnection(socket, {
    serverVersion: '16.3 (PGlite 0.2.0)',
    authMode: 'cleartextPassword',

    // Validate user credentials based on auth mode chosen
    async validateCredentials(credentials) {
      if (credentials.authMode === 'cleartextPassword') {
        const { user, password } = credentials;
        return user === 'postgres' && password === 'postgres';
      }
      return false;
    },

    async onStartup() {
      // Wait for PGlite to be ready before further processing
      await db.waitReady;
      return false;
    },

    // Hook into each client message
    async onMessage(data, { isAuthenticated }) {
      // Only forward messages to PGlite after authentication
      if (!isAuthenticated) {
        return false;
      }

      // Forward raw message to PGlite
      try {
        const [[_, responseData]] = await db.execProtocol(data);
        connection.sendData(responseData);
      } catch (err) {
        connection.sendError(err);
        connection.sendReadyForQuery();
      }
      return true;
    },
  });

  socket.on('end', () => {
    console.log('Client disconnected');
  });
});

server.listen(5432, () => {
  console.log('Server listening on port 5432');
});

You can test the connection using psql:

psql -h localhost -U postgres

You should be prompted for a password (postgres) and then brought into the psql REPL. At this point you are communicating directly with PGlite. You can verify that you are connected to PGlite by looking at the server version printed by psql:

psql (16.2, server 16.3 (PGlite 0.2.0))
...

Reverse Proxy using SNI

The server name indication (SNI) TLS extension allows clients to indicate which server hostname they intend to connect to when establishing an encrypted TLS connection. This is commonly used by HTTPS reverse proxies - without it, reverse proxies would be unable to identify which server to forward requests to since all messages are encrypted. You would need a separate IP/port pair for every server name you wish to connect to.

pg-gateway supports SNI with Postgres TLS connections to give you the same benefit. You can hook into the TLS upgrade step to retrieve the SNI server name sent by the client and use it to establish a reverse proxy connection to other Postgres servers, all over a single gateway IP/port.

In this example, clients will connect to <id>.db.example.com where id represents an arbitrary server ID we can use to look up the downstream Postgres host/port info. This implementation will terminate the TLS connection at the gateway, meaning the encrypted connections ends at the gateway and downstream data is proxied unencrypted.

We'll assume that the TLS cert used by the server has a wildcard for *.db.example.com, though if you wanted to send separate certs for each server, that is also supported (see below).

index.ts

import { readFile } from 'node:fs/promises';
import net, { connect, Socket } from 'node:net';
import { PostgresConnection, TlsOptionsCallback } from 'pg-gateway';

const tls: TlsOptionsCallback = async ({ sniServerName }) => {
  // Optionally serve different certs based on `sniServerName`
  // In this example we'll use a single wildcard cert for all servers (ie. *.db.example.com)
  return {
    key: await readFile('server-key.pem'),
    cert: await readFile('server-cert.pem'),
    ca: await readFile('ca-cert.pem'),
  };
};

// Looks up the host/port based on the ID
async function getServerById(id: string) {
  // In this example we'll hardcode to localhost port 54321
  return {
    host: 'localhost',
    port: 54321,
  };
}

const server = net.createServer((socket) => {
  const connection = new PostgresConnection(socket, {
    tls,
    // This hook occurs before startup messages are received from the client,
    // so is a good place to establish proxy connections
    async onTlsUpgrade({ tlsInfo }) {
      if (!tlsInfo) {
        connection.sendError({
          severity: 'FATAL',
          code: '08000',
          message: `ssl connection required`,
        });
        connection.socket.end();
        return;
      }

      if (!tlsInfo.sniServerName) {
        connection.sendError({
          severity: 'FATAL',
          code: '08000',
          message: `ssl sni extension required`,
        });
        connection.socket.end();
        return;
      }

      // In this example the left-most subdomain contains the server ID
      // ie. 12345.db.example.com -> 12345
      const [serverId] = tlsInfo.sniServerName.split('.');

      // Lookup the server host/port based on ID
      const serverInfo = await getServerById(serverId);

      // Establish a TCP connection to the downstream server using the above host/port
      const proxySocket = connect(serverInfo);

      // Detach from the `PostgresConnection` to prevent further buffering/processing
      const socket = connection.detach();

      // Pipe data directly between sockets
      proxySocket.pipe(socket);
      socket.pipe(proxySocket);

      proxySocket.on('end', () => socket.end());
      socket.on('end', () => proxySocket.end());

      proxySocket.on('error', (err) => socket.destroy(err));
      socket.on('error', (err) => proxySocket.destroy(err));

      proxySocket.on('close', () => socket.destroy());
      socket.on('close', () => proxySocket.destroy());
    },
  });

  socket.on('end', () => {
    console.log('Client disconnected');
  });
});

server.listen(5432, () => {
  console.log('Server listening on port 5432');
});

To test this, we can create a self-signed certificate authority (CA) and cert. In production you could use a well known CA like Let's Encrypt.

Generate the certificates using OpenSSL:

  1. Generate the CA key and certificate:

    openssl genpkey -algorithm RSA -out ca-key.pem
openssl req -new -x509 -key ca-key.pem -out ca-cert.pem -days 365 -subj "/CN=MyCA"

  2. Generate the server key and CSR (Certificate Signing Request):

    openssl genpkey -algorithm RSA -out server-key.pem
openssl req -new -key server-key.pem -out server-csr.pem -subj "/CN=*.db.example.com"

  3. Sign the server certificate with the CA certificate:

    openssl x509 -req -in server-csr.pem -CA ca-cert.pem -CAkey ca-key.pem -CAcreateserial -out server-cert.pem -days 365

Next we'll spin up a real Postgres server on localhost port 54321 using Docker:

docker run --rm -p 54321:5432 -e POSTGRES_PASSWORD=postgres postgres:16

This will act as our downstream server.

Next start the pg-gateway server:

npx tsx index.ts

Finally test the connection using psql:

psql "host=localhost port=5432 user=postgres sslmode=required"

You should be prompted for a password (postgres) and then brought into the psql REPL. At this point you are communicating with the downstream server through the reverse proxy. You can verify that you are connected to the downstream server by looking at the server version printed by psql:

psql (16.2, server 16.3 (Debian 16.3-1.pgdg120+1))
...

Note that we used localhost as the host which resulted in sniServerName being localhost instead of 12345.db.example.com. To properly test this, you will need to pass the real host name:

psql "host=12345.db.example.com port=5432 user=postgres sslmode=required"

If you wanted to test this without deploying pg-gateway to a real server, you could modify your /etc/hosts file to point 12345.db.example.com to your machine's loopback interface (acting like localhost):

/etc/hosts

# ...

127.0.0.1 12345.db.example.com

On Windows this file lives at C:\Windows\System32\Drivers\etc\hosts.

Development

npm run dev

License

MIT