We should probably move the handling out from the syncapi, but that
requires the clientapi to stream the current state which it currently
doesn't. This at least stops the sync and state handling being done in
one file.
The pre-commit hook took 45 seconds to run on my machine, which was more than
enough time for me to get distracted by a swordfight in the corridor.
Let's just run the linters (which still take 6 seconds). It's not the place of
a commit hook to run every test we can think of - that is what CI is for.
We should no longer be building anything into vendor/bin in the scripts,
so adding it to the path can lead to all sorts of confusion if old
binaries are there.
* Update gometalinter config
gometalinter now uses `maligned` instead of `aligncheck`
(https://github.com/alecthomas/gometalinter/pull/367), so we need to update our
config accordingly.
* Update gometalinter
* Disable gotype linter
gotype does not seem to play nicely with the gb vendor directory. In
particular, it wants each of our dependencies to be built and installed (see
https://github.com/golang/go/issues/10969), but (empirically) it will not
accept them being installed in `pkg` but insists on them being in `vendor/pkg`.
This presents a problem because `gb build` builds the packages into `pkg`
(which doesn't seem entirely unreasonable since `.` comes before `vendor` in
`$GOPATH`). `go install github.com/x/y` does install in `vendor/pkg` but
requires us to know the name of each package.
The general conclusion of https://github.com/alecthomas/gometalinter/issues/91
seems to have been that the easiest thing to do is to disable `gotype` for now.
* Fix `unparam` lint
* Fix goshadow lint
The motivation for this is to make it easier to see whether a travis failure is due to linting, unit tests or integration test failures, without having to look in the logs.
It also means that each job is independent, so if e.g. the linting fails then the unit tests will still be run.
This has two benefits:
1. Using channels makes it easier to time out while waiting
2. Allows us to clean up goroutines that were waiting if we timeout the
request
The way we store the partition offsets for kafka streams means that when
we start after a crash we may get the last message we processed again.
This means that we have to be careful to ensure that the processing
handles consecutive duplicates correctly.
The HTTP handlers in the components are split into reader and writer directories. This was a fairly arbitrary distinction, and turns out to not be so helpful. Most read APIs have a corresponding write API, and it is more natural for them to be in the same file rather than in different directories.
* Implement filter POSTing and GETting.
Signed-off-by: Jan Christian Grünhage <jan.christian@gruenhage.xyz>
* Add missing '}' typo introduced during merge
* Still trying to fix that merge...
* Fix linting
* Update the install instructions
To include some of the newer components.
* update INSTALL diagram
client_data goes to sync-api-server, not room-server.
Erik Johnston
Richard van der Hoff
Paul Tötterman
Andrew Morgan
Michael Telatynski
Remi Reuvekamp
Marcel
Erik Johnston
Richard van der Hoff
Mark Haines
Jan Christian Grünhage