Deploying your app to multiple regions along with your data is a great way to make your app really fast, but there are two issues:
- Read replica instances can only read from the database, they cannot write to it.
- There's an edge case where the user could write to the primary instance and then read from a replica instance before replication is finished.
The first problem is as simple as making sure you use a special fly-replay
response so Fly can pass the request to the primary instance:
But the second problem is a little harder. Here's how we visualize that:
This module comes with several utilities to help you work around these issues. Specifically, it allows you an easy way to add a special cookie to the client that identifies the client's "transaction number" which is then used by read replicas to compare to their local transaction number and force the client to wait until replication has finished if necessary (with a timeout).
Here's how we visualize that:
At the time of this writing, LiteFS just released experimental support for a
proxy server that will handle much of this stuff for you. You simply configure
the proxy server in your litefs.yml
and then you don't need to bother with the
tx number cookie or ensuring primary on non-get requests at all. The litefs-js
module is still useful for one-off situations where you're making mutations in
GET requests for example, or if you need to know more about the running
instances of your application, but for most of the use cases, you can get away
with using the proxy.
Learn more about using the proxy from this PR.
This module is distributed via npm which is bundled with node and
should be installed as one of your project's dependencies
:
npm install --save litefs-js
Unless you plan on using lower-level utilities, you'll need to set two environment variables on your server:
LITEFS_DIR
- the directory where the.primary
file is stored. This should be what you set yourfuse.dir
config to in thelitefs.yml
config.DATABASE_FILENAME
- the filename of your sqlite database. This is used to determine the location of the-pos
file which LiteFS uses to track the transaction number.INTERNAL_PORT
- the port set in the fly.toml (can be different fromPORT
if you're using the litefs proxy). This is useful for thegetInternalInstanceDomain
utility.
The best way to use this is with the latest version of LiteFS which supports the
proxy
server:
# litefs.yml
# ...
proxy:
# matches the internal_port in fly.toml
addr: ':${INTERNAL_PORT}'
target: 'localhost:${PORT}'
db: '${DATABASE_FILENAME}'
# ...
From there all you really need litefs-js
for is when you run your database
migrations. For example:
// start.js
const fs = require('fs')
const { spawn } = require('child_process')
const os = require('os')
const path = require('path')
const { getInstanceInfo } = require('litefs-js')
async function go() {
const { currentInstance, currentIsPrimary, primaryInstance } =
await getInstanceInfo()
if (currentIsPrimary) {
console.log(
`Instance (${currentInstance}) in ${process.env.FLY_REGION} is primary. Deploying migrations.`,
)
await exec('npx prisma migrate deploy')
} else {
console.log(
`Instance (${currentInstance}) in ${process.env.FLY_REGION} is not primary (the primary instance is ${primaryInstance}). Skipping migrations.`,
)
}
console.log('Starting app...')
await exec('node ./build')
}
go()
async function exec(command) {
const child = spawn(command, { shell: true, stdio: 'inherit' })
await new Promise((res, rej) => {
child.on('exit', code => {
if (code === 0) {
res()
} else {
rej()
}
})
})
}
The only other thing you need to worry about is if you ever perform a mutation as a part of a GET request. The proxy will handle forwarding non-GET requests for you automatically, but if you need to write to a non-primary in a GET, you'll need to use some of the utilities to ensure the server replays the request to the primary.
Integrating this with your existing server requires integration in two places:
- Setting the transaction number cookie on the client after mutations have finished
- Waiting for replication to finish before responding to requests
Low-level utilities are exposed, but higher level utilities are also available
for express
and remix
.
Additionally, any routes that trigger database mutations will need to ensure
they are running on the primary instance, which is where ensurePrimary
comes
in handy.
import express from 'express'
import {
getSetTxNumberMiddleware,
getTransactionalConsistencyMiddleware,
getEnsurePrimaryMiddleware,
} from 'litefs-js/express'
const app = express()
// this should appear before any middleware that mutates the database
app.use(getEnsurePrimaryMiddleware())
// this should appear before any middleware that retrieves something from the database
app.use(getTransactionalConsistencyMiddleware())
// ... other middleware that might mutate the database here
app.use(getSetTxNumberMiddleware())
// ... middleware that send the response here
The tricky bit here is that often your middleware that mutates the database is
also responsible for sending the responses, so you may need to use a lower-level
utility like setTxCookie
to set the cookie after mutations.
Until we have proper middleware support in Remix, you'll have to use the express
or other lower-level utilities. You cannot currently use this module with the
built-in Remix server because there's no way to force the server to wait before
calling your loaders. Normally, you just need to use
getTransactionalConsistencyMiddleware
in express, and then you can use
appendTxNumberCookie
as shown below.
Of course, instead of using express with
getTransactionalConsistencyMiddleware
, you could use
await handleTransactionalConsistency(request)
to the top of every loader if
you like:
// app/root.tsx (and app/routes/*.tsx... and every other loader in your app)
export function loader({ request }: DataFunctionArgs) {
await handleTransactionalConsistency(request)
// ... your loader code here
}
The same thing applies to getEnsurePrimaryMiddleware
as well. If you need or
like, you can use await ensurePrimary()
in every action
call or any
loader
s that mutate the database (of which, there should be few because you
should avoid mutations in loaders).
We're umm... really looking forward to Remix middleware...
The appendTxNumberCookie
utility should be used in the entry.server.ts
file
in both the default
export (normally people call this handleDocumentRequest
or handleRequest
) and the handleDataRequest
export.
// app/entry.server.ts
import { appendTxNumberCookie } from 'litefs-js/remix'
export default async function handleRequest(
request: Request,
responseStatusCode: number,
responseHeaders: Headers,
remixContext: EntryContext,
) {
// Most of the time, all mutations are finished by now, but just make sure
// you're finished with all mutations before this line:
await appendTxNumberCookie(request, responseHeaders)
// send the response
}
export async function handleDataRequest(
response: Response,
{ request }: Parameters<HandleDataRequestFunction>[1],
) {
// Most of the time, all mutations are finished by now, but just make sure
// you're finished with all mutations before this line:
await appendTxNumberCookie(request, response.headers)
return response
}
There are several other lower-level utilities that you can use. They allow for more customization and are documented via jsdoc. Utilities you may find helpful:
ensurePrimary
- Use this to ensure that the server that's handling the request is the primary server. This is useful if you know you need to do a mutation for that request.getInstanceInfo
- get thecurrentInstance
andprimaryInstance
hostnames from the filesystem.waitForUpToDateTxNumber
- wait for the local transaction number to match the one you give itgetTxNumber
- read the transaction number from the filesystem.getTxSetCookieHeader
- get theSet-Cookie
header value for the transaction numbercheckCookieForTransactionalConsistency
- the logic used to check the transaction number cookie for consistency and wait for replication if necessary.getAllInstances
- get all the instances of your app currently runninggetInternalInstanceDomain
- get the internal domain for the current instance so you can communicate between instances of your app (ensure you've set theINTERNAL_PORT
environment variable to what appears in yourfly.toml
).
This module uses the special .primary
directory in your Fuse filesystem to
determine the primary
(litefs primary docs), and the -pos
file to determine the transaction number
(litefs transaction number docs).
When necessary, replay requests are made by responding with a 409 status code
and a fly-replay
header
(docs on dynamic request routing).
This was built to make it much easier for people to take advantage of distributed SQLite with LiteFS on Fly.io. The bulk of the logic was extracted from kentcdodds/kentcdodds.com.
MIT