Add Redis Cache for User table #3924
Conversation
Signed-off-by: Matt Krick <matt.krick@gmail.com>
Signed-off-by: Matt Krick <matt.krick@gmail.com>
Signed-off-by: Matt Krick <matt.krick@gmail.com>
| @@ -80,7 +77,6 @@ export default class User { | |||
| this.createdAt = createdAt || now | |||
| this.picture = picture || `${AVATAR_BUCKET}/${avatarName}.png` | |||
| this.updatedAt = updatedAt || now | |||
| this.emailVerified = emailVerified || false | |||
There was a problem hiding this comment.
emailVerified lives on user.identities
| this.redisCache.prime(table, docs) | ||
| return this | ||
| } | ||
| async read<T extends keyof RethinkTypes>(table: T, id: string) { |
There was a problem hiding this comment.
read is very similar to data loader's load function.
| @@ -52,6 +53,10 @@ const threadableLoaders = [ | |||
|
|
|||
| // TODO: refactor if the interface pattern is used a total of 3 times | |||
|
|
|||
| export const users = () => { | |||
| return new ProxiedCache('User') | |||
There was a problem hiding this comment.
this lets folks still call dataLoader.get('users').load('user123'). but now, it will attempt to fetch from the cache instead of going directly to the DB
| @@ -0,0 +1,28 @@ | |||
| /* | |||
There was a problem hiding this comment.
@tianrunhe here's a general overview of the new architecture
Signed-off-by: Matt Krick <matt.krick@gmail.com>
| private callbacks = [] as {resolve: (payload: any) => void; reject: (err: any) => void}[] | ||
| private cacheHits = [] as (() => void)[] | ||
| private writes = [] as (RWrite<any> & {resolve: (payload: any) => void})[] | ||
| private ttl = ms('1h') |
There was a problem hiding this comment.
Can we make this a configuration/env value? Benefit is that for local development we can set it to a rather low value so we could test the caching locally?
There was a problem hiding this comment.
for sure, we could make the ttl an argument to the constructor, but since this is a singleton i figured it'd be just as easy to change it here, no strong preference so if you've got one feel free!
i try to keep the .env free of client variables, with the exception of 3rd party client keys like google, github, etc, since those vary based on environment.
| redis.multi(writes).exec() | ||
| return keys.map((key, idx) => { | ||
| const cachedDoc = cachedDocs[idx] | ||
| if (cachedDoc) return JSON.parse(cachedDoc) |
There was a problem hiding this comment.
Why not just return parsedDocs[idx]?
There was a problem hiding this comment.
that is a fantastic find! will fix
Signed-off-by: Matt Krick <matt.krick@gmail.com>
Today, almost every single request to the database requires looking up the User in the DB.
To minimize the hits to the DB we use a [DataLoader] (https://github.com/graphql/dataloader) inside the graphql resolvers. This reduces the number of DB hits to 1 per request, since the dataloader is disposed of after the request completes.
We can further reduce hits to the DB by caching the User in redis & busting the cache when the User is updated.
We can do better still by batching the queries together so only 1 message is sent to the DB.
We can use that same batching strategy for the cache so only 1 message is sent to redis.
We can also cache the document in memory to avoid repeatedly hitting redis & parsing the JSON.
This PR implements this batching + 3-stage cache strategy for all reads to the User table.
calling
db.readreturns the document locally, from redis, or from the DB.calling
db.writeupdates the document in the DB and updates the redis and local cache.For now, this is ONLY implemented for reads to the User table when using the userId primary key.
In the future, we can reuse this strategy for foreign keys & more types
One severe limitation: this only works with a single node instance. if node1 sends a write, its cache, redis', and the DB are all updated. however, node2's local cache will now be stale. To fix, we can use Redis' new client-side-caching, but that feature is only in v6, which DO doesn't offer yet.
TEST