In this change, we cache the zones table as part of the JS build process. This keeps the database as our source of truth, while aggressively caching the data at deploy time.
See the new README for some rationale!
I tested this by pulling up dev Honeycomb, and observing that we no longer run db queries to `zones` in the new traces for the wardrobe page. (It's a good thing we did it this way, because I noticed some code in the server that was still loading the zone anyway, and fixed it here!)
I was finding the script too slow running on my local machine, because the SQL RTTs were too slow - and with one connection, they were essentially a serial bottleneck, not taking much advantage of our concurrency.
Here, I instead add a `--dump` option, which outputs SQL to stdout. I then uploaded the resulting SQL to the DTI box, and ran it up there. Doing the network part fast on my machine, and the SQL part fast on the cloud machine!
I first considered uploading this script to the cloud machine, but it's an old Ubuntu and I couldn't figure out how to install a recent NodeJS onto it 🙃
This reverts commit 0f7ab9d10e.
The Production Vercel deploys don't seem to like how I did this build trick, even though the Preview deploys seem fine with it 🤔 Reverting for now, sent a message to Vercel support.
Just gonna bulk load all those manifests into the db, and then that should make most loads notably faster by removing the net request! 🤞
We'll still load manifests inline sometimes, but only the first time anyone pulls up the layer in impress-2020. After that, it should be cached forever!
Here's just some simple caching: we try to load the asset manifest from the db with the rest of the asset. If it's not present, we load it via HTTP, and write it to the database.
I might try to do a bulk write of manifests at some point, too.
This is because I noticed that one of the main bottlenecks in most of the endpoints now (and definitely the highest-variance) was loading from images.neopets.com.
Another approach I considered was HTTP/2 to load the manifests, because it kinda looks like the server is refusing to open all these sockets at once and effectively does the requests in waves? But images.neopets.com doesn't support HTTP/2 right now anyway, so oh well! (And that would have probably cut us down to ~250ms of HTTP time still, instead of ~600–700. Also, why is network out of Vercel so slow? :p)
I noticed that, while looking up zone data from the db is near instant when you're on the same box, it's like 300ms here!
In this change, we start downloading zone data into the build process. That way, we can have a very fast and practically-up-to-date cache (I'm not sure I've changed it in many years), while being confident that it's in sync with the database source of truth (for things like join queries).
Oops, of course, we weren't actually taking proper advantage of the dataloader here! The queries got over-complicated, but more importantly, subsequent requests to the same loader would re-submit the query!
I noticed it in the SearchPanel operation, in this Honeycomb trace:
https://ui.honeycomb.io/openneo/datasets/dress-to-impress--2020-/trace/aMuhsTjQFZY
We got bit by the "can't run anything after the response finishes" thing
so I'm just forcing the response to wait for Honeycomb submit to finish
I hope this isn't like, just awful for perf lol. but puts to honeycomb seem fast?
An even better resolution mode for "Dyeworks Pink: Peaceful Tree Garland", which doesn't leave any big chunks or holes, and instead takes all the leaves and gives them chunky outlines 😅
The "Dyeworks Pink: Peaceful Tree Garland" was a tricky case, with animated falling leaves…
we decided that having transparency in the main pet area, and some incorrect transparent holes in the trees, was a better conflict resolution for this one
Probably would be good to manually upload a Totally Good version, but like, this flag is probably good to have
Hey wow this was not so hard, just set some global styles, removed some hardcoded colors, and walked through the remaining hardcoded colors to pick a dark mode variant :) neat!!
oops another thing I missed in the refactor! I wasn't providing the selected pose to the hook, so it wasn't styling any of the poses as selected. Now it does again!
Oops, I removed some fields from the pose object, so the aria-label for the options were coming out as "undefined and undefined"!
Now they come out correct!
I tried to get the alignment juuust right in Chrome and Firefox, but Dice's computer was still including a 1px sliver of border.
In this change, I'm adding 1px of extra padding between the border and the Flash area. I figure that, if the screenshots are incorrectly aligned by 1px, that's no big deal and I'm okay accepting that slightly-incorrect upload rather than having folks not be able to use the tool!
I skipped this in the past runs because I had a hard time getting consistency from the results… but they seem to be behaving now?
It really seemed like there were some races on certain query orders… maybe there still is, but my more-reliable connection today is making them resolve in a more consistent order?
Anyway if I see goofs again, I'll consider adding a snapshot matcher that isn't picky about query order 🤔
I figure we can get even better messages than "Access denied" in case this happens again!
Also, I need a new deployment to trigger the new env vars, so why not make it count? :)
My hypothesis is that this missing `await` is causing prod Vercel to cut off the request before the async upload finishes. Omitting this `await` was a mistake, so I'm just adding it right in!
I also add logging for the success case. That way, if this isn't actually a fix, I can at least watch the logs and confirm whether the app thinks it's getting to here.
This happened in an earlier change too, I forget where, but there's some compiler bug that causes comments placed in this kind of spot to yield this error in production, but not in dev:
https://reactjs.org/docs/error-decoder.html/?invariant=152&args
Oh well, moved the comment!
On mobile, it was pretty annoying that you had to show the controls by tapping the preview area to simulate a hover—because it could also click the underlying elements, and do bad stuff!
Here, I add a click capture to prevent the clicks from going down if the controls aren't visible. And I add a toggle, so that you can dismiss the controls, like how YouTube feels :)
Huh, weird how we seem to need preventDefault for buttons, but stopPropagation for links? idk, in any case, the Remove button was doing that thing where it clicks Remove but also bubbles up to the container and clicks it too :p
This was a surprisingly big win! Item is heavier than it looks, because it has like 6 Chakra components, which aren't expensive but aren't _cheap_ in a re-rendered list that needs to be fast, you know?
And it's even more important on search, where there's a lot of items on the page. (we should virtualize it too but that's a thing for another day)
I noticed that item wear/unwear is slow on mobile, because we re-render the whole app tree, and my laptop handles that super fine, but my few-years-old fun takes ~300ms, which is very noticeable.
There's some hacks we could do to get faster feedback, but first I'm diving into the render tree to find the unnecessary renders and stop 'em! That should help build perf across the board, rather than in just one spot, and hopefully be less of a weird sore spot :)
The button section was capable of flex-shrinking, and having those wrap pretty much always looks bad imo. Here, we get more precise about the areas and their flex rules, including only letting the name area grow/shrink. (If the screen is too small for the name to wrap further, the panel area container gains a horizontal scrollbar, which feels like a really good compromise imo)
