\n```\n\n## Acknowledgements\n\nSpecial thanks to the authors of the `pglite` project, without which our journey of enabling Lantern in the web would take a lot longer.\n\nWe would like to thank the authors of emscripten as well, for building a great library and for being very responsive when we reported an issue.\n\nThe Wasm debugging [tutorial](https://developer.chrome.com/blog/wasm-debugging-2020/) by Ingvar Stepanyan was very helpful for building this demo as well. Consider listening through it when hacking with Wasm!"}

Postgres

Dynamically loaded extensions in Postgres in the browser

July 12, 2024 · 6 min read

At a recent AGI House hackathon in San Francisco, we set out to run our Postgres vector extension directly in the browser. This would allow our documentation examples and small demos to be fully executable within the browser using a Postgres instance compiled to Wasm and running on the client browser.

We built on top of pglite, which uses a modified version of Postgres that enables running Postgres in the browser and NodeJS. We added support for dynamically loading extensions into pglite, and built an application demonstrating vector search inside the browser via dynamically loaded lantern and pgvector extensions.

Head over to our demo app to interact with a fully functional Postgres running inside the browser - The demo showcases geospatial search using vector indexes from Lantern. You can dynamically load extensions into a fully in-browser Wasm-compiled Postgres server, interact with the database via the Postgres psql shell, and more!

Goal

With this project, we had the following goals:

  • Do vector search from inside the browser - This requires loading the lantern extension
  • Avoid loading unused extensions - We should only load an extension when it is actually necessary / included
  • Load extensions dynamically, from a separate compile unit - This ensures extensions can be compiled and added independently

Working Demo and Tested Extensions

Here’s a quick demo video showcasing vector search, geospatial search, and encryption using Postgres extensions.

Source code:

We have compiled and tested the following extensions:

Technical Enhancements

We relied heavily on the pglite project to get a basic Postgres setup running inside the browser. However, the project did not support dynamically loading extensions or compilation support for the specific extensions we had in mind.

Below, we describe the more interesting hurdles we had to overcome to achieve our goals, building off of pglite.

Dynamic Linking Enabled

Postgres loads extensions dynamically via dlopen. However, dlopen is not available via emscripten when the the program is not compiled as a main module (MAIN_MODULE=1).

When compiling Postgres with MAIN_MODULE=1 which also exports all the symbols defined in binary, we ran into a conflict with the proc_exit symbol, reserved by the WASI runtime.

Postgres Compilation Error

This was a bug in emscripten and we have opened an issue to follow up with a proper fix.

Meanwhile renaming the offending symbol in Postgres source to pg_proc_exit resolved the issue, though it may break extensions that use the proc_exit function.

Since emscripten implementation of dlopen requires the fetch API in browsers, we needed to add the ASYNCIFY flag to wrap the code in an asynchronous runtime. As a result, we also needed to call the Postgres functions from JavaScript via ccall.

Diff to add ccall

Function Pointer Casts Fix

We needed to remove the EMULATE_FUNCTION_POINTER_CASTS flag, so we could dynamically resolve symbols using dlsym.

Simply disabling it made the application crash since Postgres used dynamic function pointer casts in one instance of a query parse tree walker function with wrong arguments (source)

Postgres Null Error

After some research we found a patch in Postgres upstream which resolved the issue in later Postgres versions. So, just backporting the patch allowed us to drop the EMULATE_FUNCTION_POINTER_CASTS flag.

Dynamically Loading Extensions

Allow loading non-existent web paths

When loading an extension, Postgres wants to ensure the library file is already on the file system, but this is not the case when we load the extension via a URL. Commenting the relevant check in Postgres source code prevents the failure until we are able to inject the library file into the file system as discussed later.

Diff to allow web paths

The module_pathname in the control file should contain the full URL or path to the extension Wasm file, but this will only be loaded when the CREATE EXTENSION statement is executed.

Control file example:

plaintext

comment = 'vector data type and ivfflat and hnsw access methods'
default_version = '0.7.2'
module_pathname = 'https://storage.googleapis.com/lantern-wasm/pgvector/vector.wasm'
relocatable = true

We also needed to modify the locateFile function in pglite to handle URLs for extension Wasm files correctly.

Inject dynamic extension into WASM file system

We adapted an existing extensions interface proposal (PR) in pglite, and implemented necessary backend support to enable dynamically loading extensions.

The property pathOrUrl should be a URL or path in a filesystem containing a control file, SQL script, and Wasm library for a specified extension.

We added the addExtension(name: string, extension: Extension) function for registering extension control and SQL files after database initialization. With this API, it is not required to register all extensions beforehand, but they can also be registered via user interaction after the database is running (this is demonstrated in the demo app for the pgcrypto extension).

Alternatively, if we know an extension will be necessary on startup, we can load it as the database gets initialized. On startup we will concurrently fetch control and corresponding SQL files for each of the specified extension and put them under the $sharedir in emscripten module filesystem. Then later the CREATE EXTENSION command can be used to fetch the Wasm file specified in module_pathname and dynamically load it.

javascript

<script type="module">
import { PGlite } from "../dist/index.js";

// In-memory database:
const pg = new PGlite({
    extensions: {
        lantern: {
            // this path should contain lantern.control, lantern--$version.sql and lantern.wasm files
            // the module_pathname in control file will be
            // https://storage.googleapis.com/lantern-wasm/lantern-dev/lantern.wasm
            pathOrUrl: 'https://storage.googleapis.com/lantern-wasm/lantern-dev',
            setup: () => null,
        },
        vector: {
            pathOrUrl: 'https://storage.googleapis.com/lantern-wasm/pgvector',
            setup: () => null
        },
    }
});

await pg.waitReady;

// We can create the extensions specified in startup configuration
await pg.exec(`CREATE EXTENSION IF NOT EXISTS lantern;`);
await pg.exec(`CREATE EXTENSION IF NOT EXISTS vector;`);

// We can also load extension after DB has been started
await window.pg.addExtension('pgcrypto', {
    pathOrUrl: 'https://storage.googleapis.com/lantern-wasm/pgcrypto',
    setup: () => null
});
await pg.exec(`CREATE EXTENSION IF NOT EXISTS pgcrypto;`);
</script>

Acknowledgements

Special thanks to the authors of the pglite project, without which our journey of enabling Lantern in the web would take a lot longer.

We would like to thank the authors of emscripten as well, for building a great library and for being very responsive when we reported an issue.

The Wasm debugging tutorial by Ingvar Stepanyan was very helpful for building this demo as well. Consider listening through it when hacking with Wasm!

Authors

Varik Matevosyan

Varik Matevosyan

Software Engineer

Share this post