Apparently 'my sub' captured the $conn variable and held a ref on it
even beyond the parent sub scope. 'my $x = sub {}' doesn't do that.
Getting the ref counts right is important here for the last test to work.
(Found while I was inspecting the refcount effects of the new ->conn()
methods with Devel::Peek)
Turns out VNDB has a few places where request data is directly used for
column names in VALUES/SET/WHERE clauses. These are already restricted
to known strings through the use of FU::Validate, but an extra layer of
protection seems warranted here.
And use it for automatic output compression in FU, as (potentially)
faster alternative to Compress::Raw::Zlib.
Was also planning to maybe add support for Zstd or Brotli, but given the
performance of libdeflate, I'm not sure that's really necessary. Brotli
does tend to do a better job at compressing HTML, though.
I'm breaking stuff left and right while I still can.
Idea: "key_names" validation?
Idea: "tuple" validation that works like "keys" but for arrays.
(i.e. { tuple => { $index => $schema } }, could make "missing" and
"unknown" work for arrays, too)
This allows all built-in options to be duplicated inside a single
schema, the semantics of which are the same as the kind of merging
done as part of inheriting options from custom validations.
This also causes all 'keys' and 'values' validation schemas to be
merged, which changes error messages a bit but is great for
introspection. Probably slightly improves performance as well.
API is not super convenient and implementation is lousy, but uploading
files is not a super common operation so that should be fine.
At least it supports large files with only a single in-memory copy.
Using a consistent numeric interpretation for timestamps, dates and
times simplifies a bunch of operations. It certainly simplifies
conversion between the Postgres formats.
Copied from TUWF::Validate with a few small changes. I have a few more
features planned, but let's see how this goes first.
It's been an incredibly useful module in the past, I'm not sure right
now if I had ideas for potential improvements at some point, will need
to check notes.
And with this, I have a working rewrite of the manned.org backend into
FU. \o/
The $st->row methods are very useful even for queries that may not
return anything, so their old behavior was unhelpful. Interestingly
enough, the error-on-multiple-rows did catch an actual bug in
Manned.org, so I'm keeping that behavior.
...I was hoping not to have to implement the date type, because date
conversions suck, but it turns out manned.org actually needs it.
(Only to then convert it into a Unix timestamp again, hmm, maybe this
string conversion isn't useful at all?)
What I'd really like, in addition to this, is a way to extract a query
from an $st object that can be run in the psql CLI. VNDB has a debugging
feature for that, but it's less trivial to make that work with binary
query parameters.
And remove UTF-8 check in JSON writer. It honestly feels kind of silly
to do that validation there while I've never done similar validations in
any other output routines - including this XML writer.
FU::XMLWriter is a copy of TUWF::XMLXS with a bunch of improvements
applied: now uses refcounts to determine the current output instance,
auto-generates XS functions and has faster escaped string output -
inspired by the JSON writer.
TODO:
- Integrate into FU
- Do something with bool attribute values
- Benchmarks
- Should $content be optional for all tags? The reason they weren't in
TUWF::XMLXS is because TUWF::XML supports opening tags without closing
them, but that idea turned out to suck and isn't supported anymore.
This is hopefully the last XS module for the FU framework. The only C
code being written now should be bug fixes and extending FU::Pg with
some planned features. Already ended up with more C than I had
planned...
I initially planned to only implement the bare minimum to support
FastCGI under nginx, but ended up implementing the full protocol
instead. This is more code than I had expected and the code is also less
trivial than I had hoped. Will need to do more testing, pretty sure
there's bugs left.
Also TODO: test under alternative process managers + document
FU_LISTEN_PROTO.
I've also removed the max_request_body setting, this is something that
really ought to be configured in the web server instead.
