We used get_multi when preparing the proxies to decide which to
load from the database, but then sent multiple get requests to
Memcache to re-fetch the same data from that get_multi. Silly!
Use the data that's already stored on the proxy anyway.
Right now we're spending too much time expiring cache keys when
getting contributions. The longer-term fix is to move it to a
background task, but it's good to restrict deletions only to usable
locales rather than all the ones that Rails theoretically supports.
Fun bug! If you edit an outfit, but the outfit loads before the
closet items do, then we clone the outfit to give it its new
identity and therefore forget about its item load callbacks.
Now we have a cheap hack to forward item load data to the
outfit's clones. Hooray! Hope this doesn't break tons of things!
That is, Neopets.com will raise an error when you try to equip a
Kyrii Mage Cape to a pet who's already wearing Ceremonial Shenkuu
Warrior Armour, since the armor restricts the Collar zone which
the cape occupies. DTI, however, would just hide the Collar zone,
as if it were biology. Now, however, DTI will unwear the armor
when you wear the cape, and vice-versa (despite the restriction
relationship being one-directional).
Some lame benchmarking on my box, dev, cache classes, many items:
No proxies:
Fresh JSON: 175, 90, 90, 93, 82, 88, 158, 150, 85, 167 = 117.8
Cached JSON: (none)
Fresh HTML: 371, 327, 355, 328, 322, 346 = 341.5
Cached HTML: 173, 123, 175, 187, 171, 179 = 168
Proxies:
Fresh JSON: 175, 183, 269, 219, 195, 178 = 203.17
Cached JSON: 88, 70, 89, 162, 80, 77 = 94.3
Fresh HTML: 494, 381, 350, 334, 451, 372 = 397
Cached HTML: 176, 170, 104, 101, 111, 116 = 129.7
So, overhead is significant, but the gains when cached (and that should be
all the time, since we currently have 0 evictions) are definitely worth
it. Worth pushing, and probably putting some future effort into reducing
overhead.
On production (again, lame), items#index was consistently averaging
73-74ms when super healthy, and 82ms when pets#index was being louder
than usual. For reference is all. This will probably perform
significantly worse at first (in JSON, anyway, since HTML is already
mostly cached), so it might be worth briefly warming the cache after
pushing.
That is, once we get our list of IDs from the search engine, only
fetch records whose JSON we don't already have cached.
It's simpler here to use as_json, but it'd probably be even faster
if I figure out how to serve a plain JSON string from a Rails
controller. In the meantime, requests of entirely cached items
are coming in at about 85ms on average on my box (dev, cache
classes, many items), about 10ms better than the last
iteration.
Specifically, we were running a find_or_initialize_by for all 50
hangers, which isn't great. Collation logic is more complicated this
way, but query count is way lower.
Additionally, compare against hanger.list_id instead of hanger.list,
because hanger.list will fire a query if list_id is non-nil, but that
nil ID tells us everything we needed to know, anyway.
Bug report that this resolves:
...However, when I was using the "Import from SDB" tool just a few
minutes ago, it ended up adding EVERY neocash item into the "Not
In A List" section, regardless if I already that item imported
into my "Your Items". So, basically.. I had duplicates of
everything and it would not allow me to move them around into
separate catergories or anything. I know that every other time i've
used the import tool, it would only add NEW items that are not
currently already in my lists yet.
Most of the reasoning is documented in the big comment. In short, we tried
to solve the problem with caching, but the caching should hardly be necessary
now that the bottleneck should be fixed. We'll see on production if it
actually solves the whole problem, but I've confirmed in the console that
redefining this function makes random_basic_per_species (as called during
rendering) a ton faster. And this way we keep our randomness, woo!
This is a surprisingly huge performance gain. On my testing (with
cache_classes set to true to also cache templates), this sped up
closet_hangers#index rendering by a factor of 2 when there were a
significant number of items. Cool beans.
I think we can even hold off on the individual hanger caching now:
we've made the closet hanger partial tons faster by moving forms out
of them and doing this cache check earlier. I'm expecting significant
performance gains both here and on items#index (though less so there).
I'll deploy and see how much it helps in production; if not enough, we
can look at the layered caching of hangers, lists, groups, full pages,
etc.
So glad we don't *have* to move to a pagination model!
We lose no-JS support, which I kinda miss, but caching is gonna be more
important down the line. Delete form moves next, then we cache.
CSRF token changes: it looks like, by setting a data attribute in AJAX, I
was overwriting the CSRF token. I don't remember it working that way, but
now we use beforeSend to add the X-CSRF-Token header instead, which is nicer,
anyway. The issue might've been something else, but this worked :/
The CSS was also not showing the loading ellipsis properly. I think that's a
dev-only issue in how live assets are being served versus static assets, but
may as well add UTF-8 charset directives everywhere, anyway.
items#show has been very slow recently, and I think it's because there's a lot
of querying to be done. Another option would have been to attempt to
short-circuit Item#supported_species if not body specific, but that would
still leave us with 1s load times for body specific items, which is not
satisfactory. The short-circuiting might still be worth doing, but probably
not now.
I'm also not sure that this is actually the core performance problem, but
we'll see. It definitely helped on the dev server: items#show took about
200ms on item pages where everything but species images were cached, then
took about 30ms on subsequent loads. Looking like a good candidate.
TNT has started serving half-removed Corridor of Chance effects:
it has the asset ID and URL and all, but the zone ID is blank.
RocketAMF has patched the empty key bug, and now we ignore assets
associated with empty keys.
Specifically, the Tyrannian Meerca Spear is a pb item that contains
"pea", so its item page is only willing to show a Pea Chia. Now,
a color must be a whole word in the item name for special color
determination to work.
A few key changes:
* Don't reload the whole pet 8 times!! Sooo many bad things
happen, including redundant lookups of everything else and
too many item saves and reindexes. Instead, fetch the item
data, apply it to the items, and then save the items (once
each!)
* Updated my branch of globalize3 to be even better at avoiding
redundant queries when saving. Woo.
* Last realization: wrapping all the item saves in a single
transaction works wonders. COMMIT seems to have high overhead,
so doing only one took it from 50ms * 10 or whatever to 60ms.
Good stuff.
We were joining to the translations table to sort records
alphabetically, but then it sorted by *all* of the translations in
some strange way. Now use with_translations to restrict the join
to the current locale.
In particular, pet#load was handling locale-switching itself, but wasn't
switching back to original locale on error. We could've used a rescue
block, but, when I18n.with_locale is so cool, may as well use it fully.
We originally had a regression on name-matching, where, among
other issues, `straw hat` returned items containing both "straw"
and "hat", which isn't really helpful behavior since we're sorting
alphabetically. Now, `straw hat` behaves as expected.
Additionally, "phrases like these" behave as expected, too.
Confirmed features:
* Output (retrieval, sorting, etc.)
* Name (positive and negative, but new behavior)
* Flags (positive and negative)
Planned features:
* users:owns, user:wants
Known issues:
* Sets are broken
* Don't render properly
* Shouldn't actually be done as joined sets, anyway, since
we actually want (set1_zone1 OR set1_zone2) AND
(set2_zone1 OR set2_zone2), which will require breaking
it into multiple terms queries.
* Name has regressed: ignores phrases, doesn't require *all*
words. While we're breaking sets into multiple queries,
maybe we'll do something similar for name. In fact, we
really kinda have to if we're gonna keep sorting by name,
since "straw hat" returns all hats. Eww.
For example, the Meerca Maid Tray is a foreground item, so the SWF is marked
as compatible with all body types, but the item itself is clearly marked as
Meercas-only. items#show reflected this properly, but the swf_assets#index
call that the wardrobe uses ignored item.species_support_ids.
So, /bodies/:body_id/swf_assets.json?item_ids[]=... was deprecated in favor
of /pet_types/:pet_type_id/items/swf_assets.json?item_ids=[]..., which is
much like the former route but, before loading assets, also loads the pet
type and items, then filters the items by compatibility, then only loads
assets for the compatible items.
This one was actually pretty darn clever - nobody's abused it, but
I was reading a blog post where someone described this type of
issue, I realized it was a brilliant attack, and then realized
DTI was vulnerable. Oops. Thanks for the solution, Jamie!
http://jamie-wong.com/2012/08/22/what-i-did-at-khan-academy/#XSS+Fix
Many forms on the site contain a hidden authenticity_token field,
unique to each visitory. If a user submits a request with an
invalid authenticity_token, Rails assumes that it's a CSRF attempt
and logs out the user. So, if we happen to cache those forms with
authenticity_token fields, all users who use that form will have
the same authenticity_token (valid for only the first user who
saw the form, invalid for everyone else), and all requests made
through that form will log out the user. Bad news.
So, we stopped caching those forms. Yay!
Use the ImageMagick flatten command to generate the output all at
once instead of compositing each layer individually, and download
the layers in parallel. On my box, saving roopal27 five times took
a total of 30 seconds before, whereas now it takes 7 seconds. I
expect it to be even better on the production box, where latency
is even lower.
