tebello-thejane / bitx.hs Goto Github PK

Currently a new connection manager is created for every request. This is fairly expensive, mainly because for TLS all of the certificates in the certificate store need to be parsed and loaded. If I replace globalManager with this, my CPU usage goes down a lot:

{-# NOINLINE yuck #-}
yuck :: NetCon.Manager
yuck = unsafePerformIO (NetCon.newManager NetCon.tlsManagerSettings)

globalManager :: IO NetCon.Manager
globalManager = pure yuck

(Obviously this isn't a great thing to do, it's just a quick'n'dirty PoC of sharing the manager)

What we probably want is some way to pass a manager in to the operation to be used, but I'm not quite sure how to do this without creating duplicate variants of every operation in the API, which sounds tedious.

Currently these are Int, which is fine for XBTZAR since no decimals are allowed in the price, but useless for ETHXBT since the prices are always fractional in this market.

The travis jobs randomly time out when testing live API connectivity. This might be due to the default (non-container-based) infrastructure sharing the same IP address between multiple instances running the same test, and thus flooding BitX.

Perhaps moving to container-based infrastructure might help.

Luno recently added this (along with the ETH asset). Should be a pretty easy change, I'll try send a PR shortly.

If you request trade information where no trades match, the Luno API returns null instead of an empty array, causing a parse failure.

Citing from http://hydra.cryp.to/build/926883/nixlog/1/raw:

src/Network/Bitcoin/BitX/Internal.hs:25:8:
    Could not find module ‘Record.Lens’
    Use -v to see a list of the files searched for.

λ> B.postEncode $ B.OrderRequest B.XRPXBT B.BID 673 0.00002683
[("pair","XRPXBT"),("type","BID"),("volume","673"),("price","0.000026")]

Not sure where we're losing precision but this is clearly wrong!

I have some code implementing the streaming websocket API which I'd like to contribute, but it probably needs some more cleaning up. Also I used type-level symbols for pairs rather than value-level constants, so the API will probably need to be changed. Filing this so I hopefully don't forget forever.

We have basically given up on Travis-CI as it has a lot of noisy failures.

• Are we trying to support too many GHC versions?
• And what about the jobs which just decide to fail on whim?

Get this shit stable and fixed, so that we don't ignore valid failures due to all this noise.