jhthorsen
commented
2 years ago Thanks for creating the issue, though I did misunderstand your request. I thought you wanted to say "turn json decoding on for all messages, instead of a specific channel". I think I misunderstood.
Here are some comments:
...all handlers running within the same UNIX process...
That is not correct. It does so for all topics in the given $pubsub object. It sounds like what you are doing is that you create one $redis object for the application and then you do $redis->pubsub->json("foo") which will try to encode/decode all "foo" messages.
So the results on multi-process servers (prefork or hypnotoad) differ from what can be seen in a single-process development environment (morbo), which is a bit counterintuitive
There really isn't any difference. There's no globals, even though you might use a singletong $redis and $redis->pubsub object in your app. (Note: pubsub() is an attribute, not a method returning a fresh pubsub)
I think there should not be a standalone ...
That's not going to happen, unless you change https://metacpan.org/pod/Mojo::Pg::PubSub#json as well. The idea is that you should be able to swap between different Mojo::Whatever::PubSub libraries without any cognitive overhead.
Suggested solution:
Since $redis->pubsub is an attribute and it sounds like you are using a "global" $redis helper/object, then I suggest something like this:
use Mojolicious::Lite;
use Mojo::Redis;
my $redis = Mojo::Redis->new;
# Make sure you enable json for all json-channels up front, and tell you developer to never call json() inside an action
$redis->pubsub->json($_) for qw(admin all foo bar);
helper redis => sub { $redis };
Yenya
Grinnz
There is a problem with
$pubsub->json($channel)->...call chain - it switches the given channel to a JSON en/decapsulation, and does so permanently. But because all requests in a single server process share the same pubsub connection, the->json()call switches the channel to JSON for all handlers running within the same UNIX process, even though they did not call->json()themselves. So the results on multi-process servers (prefork or hypnotoad) differ from what can be seen in a single-process development environment (morbo), which is a bit counterintuitive.Our use case, as requested on IRC: we have a websocket app with two types of users, say teacher(s) and students. The app uses two pubsub channels: "
all" for every user, and "admin" for one-way communication of users with the teacher (think answering the teacher's questions). So the teacher's websocket handler subscribes both toallandadminchannels, and switches them to JSON mode. The students' handlers only subscribe toall, also with->json('all')but ocasionally send messages also toadmin, even though they do not listen on that channel. The result depends whether the student's handler runs inside a process where also a teacher's handler is currently running. If it does, the handler's->notify()calls send the hashrefs as encoded JSON. Otherwise, the message gets stringified to something likeHASH(0xdeadbeef). So the student's handler have to either call->json('admin')->notify('admin')every time, or track the channels which were already switched to JSON mode by itself.I can imagine that for debugging purposes, somebody would even want to write a listener for raw messages on some channel, even though it is running inside the same process as other handlers that expect JSON messages on that channel. So switching a particular pubsub connection to the JSON mode globally for a channel is not correct.
I think there should not be a standalone
->json(), but instead something like$pubsub->notify_json()and$pubsub->listen_json(), and the->json($channel)->listen/notify($channel)syntax should be deprecated.What do you think about it? Thanks,
-Yenya