Bencoded dictionaries' keys are not always sorted about chihaya

[jzelinskie]

We've known about this for since roughly around the time chihaya/bencode#2 was posted.

Bencode is performance-critical to the HTTP frontend, so I'm hesitant to add the sort unless you know of broken clients/behavior that are caused by the nondeterminism. This focus on performance and cutting out unnecessary functionality is why we use our own bencode implementation.

[cypheon]

Hi, thanks for the quick response! It's good to hear you're aware of the issue, unfortunately it did not turn up in my quick search.

I initially noticed this, because BitTornado (the Python library) was crashing/throwing exceptions regarding invalid bencoded data, so there's at least one client having trouble with keys not being in-order.

I honestly didn't really consider performance while writing the fix, since most (all) of the dicts in often-occurring messages (i.e. responses to announce messages) only have in the order of magnitude about 4 or 5 keys (interval, min_interval, peers). But then again, I didn't do any benchmarks, so it's just a hunch.

All in all, I think complying to the spec is in itself valuable. If this impacts performance too much, maybe the interface could be switched from a "Dict" to a slice of (Key, Value) tuples? That would move the responsibility of sorting the keys to the part of code where the response is generated. Of course that would require a much bigger change in the code.

[pushrax]

There are some cases where the dict can potentially be very large

chihaya/frontend/http/writer.go

Lines 75 to 86 in e7f43ee

	filesDict := bencode.NewDict()
	for _, scrape := range resp.Files {
	filesDict[string(scrape.InfoHash[:])] = bencode.Dict{
	"complete": scrape.Complete,
	"incomplete": scrape.Incomplete,
	}
	}
	return bencode.NewEncoder(w).Encode(bencode.Dict{
	"files": filesDict,
	})
	}

I agree that the best way would be to construct the dicts in order, to avoid creating another copy of keys and sorting later. However, I also agree that in practise it's probably not much of a problem to just do the sort when marshalling. A benchmark would definitely be the important deciding factor here.

[Aranjedeath]

I am running this patch in production and we will see if there is some performance issue with this or not.

[Aranjedeath]

Performance of this patch is not making me scream, so if it is unoptimized... the optimized solution will be fine.