This is easier to read than having to understand the arguments 1…n of
the cache decorator are ‘magic’, and gives us more flexibility about
how the cache keys are formatted, eg being able to add words in the
middle of them.
Also changes the key format for all templates to be
`service-{service_id}-templates` instead of `templates-{service_id}`
because then it’s clearer what the ID represents.
A lot of the frequently-used pages in the admin app rely on the API to
get templates.
So this commit adds three new caches:
- a single template version (including a key without a version number,
which is the current version)
- all the templates for a service
- all versions of a template
The first will be the most crucial for performance, but there’s not much
cost to adding the other two.
`@cache.delete('user', 'user_id')` is easier to read and understand than
`@cache.delete('user', key_from_args=[1])`. This will become even more
apparent if we have to start doing stuff like `key_from_args=[1, 5]`,
which is a lot more opaque than just saying
`'service_id', 'template_id'`.
It does make the implementation a bit more complex, but I’m not too
worried about that because:
- the tests are solid
- it’s nicely encapsulated
Most of the time spent by the admin app to generate a page is spent
waiting for the API. This is slow for three reasons:
1. Talking to the API means going out to the internet, then through
nginx, the Flask app, SQLAlchemy, down to the database, and then
serialising the result to JSON and making it into a HTTP response
2. Each call to the API is synchronous, therefore if a page needs 3 API
calls to render then the second API call won’t be made until the
first has finished, and the third won’t start until the second has
finished
3. Every request for a service page in the admin app makes a minimum
of two requests to the API (`GET /service/…` and `GET /user/…`)
Hitting the database will always be the slowest part of an app like
Notify. But this slowness is exacerbated by 2. and 3. Conversely every
speedup made to 1. is multiplied by 2. and 3.
So this pull request aims to make 1. a _lot_ faster by taking nginx,
Flask, SQLAlchemy and the database out of the equation. It replaces them
with Redis, which as an in-memory key/value store is a lot faster than
Postgres. There is still the overhead of going across the network to
talk to Redis, but the net improvement is vast.
This commit only caches the `GET /service` response, but is written in
such a way that we can easily expand to caching other responses down the
line.
The tradeoff here is that our code is more complex, and we risk
introducing edge cases where a cache becomes stale. The mitigations
against this are:
- invalidating all caches after 24h so a stale cache doesn’t remain
around indefinitely
- being careful when we add new stuff to the service response
---
Some indicative numbers, based on:
- `GET http://localhost:6012/services/<service_id>/template/<template_id>`
- with the admin app running locally
- talking to Redis running locally
- also talking to the API running locally, itself talking to a local
Postgres instance
- times measured with Chrome web inspector, average of 10 requests
╲ | No cache | Cache service | Cache service and user | Cache service, user and template
-- | -- | -- | -- | --
**Request time** | 136ms | 97ms | 73ms | 37ms
**Improvement** | 0% | 41% | 88% | 265%
---
Estimates of how much storage this requires:
- Services: 1,942 on production × 2kb = 4Mb
- Users: 4,534 on production × 2kb = 9Mb
- Templates: 7,079 on production × 4kb = 28Mb
