Remove some assumptions that prevent subscription listeners from working on an
instance. This means allowing the host to forward the event from the instance
to the controller; correctly storing subscription listeners by address rather
than by connection; and a bunch of network and event stuff to correctly clean up
a listener.
Remove the assumption that all subscriptions events contains the updates
property. This meant fixing inconsistent types; always checking if the property
existed; fixing lots of the tests that made incorrect assumptions.
Minior fixes:
- Fix double firing of listener callbacks on first connection.
- Fix synced being set to true too early on first connection.
- Fix edge case interaction between last response and last snapshot for values
with updated set as 0.
Also change the the output made during the start of tests to indicate something
is happening when start up is slow and allow for silent testing to reduce the
amount of text output to the console.
Use a separate folder for the TypeScript build for Node.js in the dist
folder. This makes each type of build consistently have its own output
folder instead of being interleaved with the Node.js build.
If a link is closed while the prepareDisconnect action is in progress
the link will still try to signal the disconnect ready over the socket
when the prepareDisconnect action completes, which throws due to the
socket no longer being present. This error is not caught due to the
catch happening before the finally on the promise and triggers an
unhandled rejection.
Fix by rearranging the finally to come before the catch on the promise
so that errors are always handled, and checking that the link still has
a session before attempting to send the disconnect ready signal.
When messages were forwarded they were sent using the low level
connector .send method which accepts a message and sends it over the
link. This method sends the message as is with no modification which
means that at no point during the forwarding logic was the message
sequence number that originated from Link A renumbered to the correct
sequence number to send out on Link B which it was forwarded on.
This in turn caused the logic which drops acked messages from the send
buffer to break which in turn meant the resume could cause duplication
of message and/or messages to be lost. With enough duplicated messages
the resume of one connection could cause other connections to drop and
also initiate resume.
Fix by renumbering the sequence on forwarded messages. In the future
the sequences sent and received should also be checked to make sure this
kind of problem does not happen again, this is not done here because
partially updating a running cluster with sequence problems to one that
checks and enforces sequences numbers will be problematic.
Use the JSON data class pattern for User, Role, and PlayerStats and move
them into lib/src/data. This means that User is purely data now and
can't reference role objects directly. To make it possible easily check
permissions attached to roles and store data only visible on the
controller on user objects a ControllerUser type is introduced which is
linked with the UserManager and can therefore reference role objects
directly.
Resolves#525.
Only the controller has access to the secret needed to sign user tokens
so it makes sense for the user signing code to live in the controller.
This saves about 20ms on the time ctl takes to start up due to lib no
longer needing to import jsonwebtoken.
Move all functions and classes currently available through lib/name and
libName to be under a single flat namespace in lib. This make
reorganising the lib namespace much simpler as the location of symbols
can move without breaking the consumers of lib.
The multi-entrypoint design was already problematic for the web plugin
system which couldn't effectively use it due to how webpack's shared
module system is awkward to set up with multiple entrypoints per package.
Split up the action of registring an event or request data class to be
decoded by the Link class and the action of registring a handler for
those requests and events. This moves the validation work from Link
instantiation time to startup initialisation which makes logic errors
fail on startup instead of failing instantiation of the link when a
connection attempt is made.
Additionally implement a partial validation logic for messages received
using the newly refactored message registration. This unfortunately
ended up merged with the registration refactor.
Reverse the direction of the API for sending messages from
messageDefinition.send(link, { content });
to
link.send(new MessageClass(content));
and convert all message definitions to use the JSON serialisable class
pattern and explicit registration of message handlers, along with a
end-to-end based routing of messages instead of link to link.
The experience I gained from the existing way messages were defined and
sent over links was that it was difficult to understand for people
unfamiliar with the system. The automatic registration of message
handlers ment that code was being called as if by magic from nowhere
with no easy way to discover it. The manual forwarding and rules for
what can be sent where and how obscured its working.
The use of pure JSON as both input and output from message handlers made
it tedious to work with. Data had to be manually converted between the
program representation and the over the wire format, which made handler
code unplesant to write. Particularly because the over the wire format
used snake_case while the program code uses camelCase for variables.
The idea behind the decision to do it this way was to make it more clear
which data came from the wire and which from the program, and while this
worked it didn't make message handlers any nicer to write.
The direction of the API and magic registration of message handlers also
meant that a future migration to TypeScript would be troublesome. While
it may be possible to implement, the degree of TypeScript magic required
would go starkly against the principle of keep it simple stupid.
By refactoring to using the JSON serialisable class pattern most of
these issues go away. There is no longer a need to distinguish the on
wire format of messages with the program representation because the only
place that deals with the wire format of messages are the toJSON and
fromJSON methods on the classes themselves. Message handlers will work
with proper instances of classes instead of raw data, something that
makes them easier to write, and trivial to type with TypeScript.
End-to-end routing also simplifies sending messages that need to do
mulitple hops to reach the destination. And removes the need for having
all messages be specified in plugin's info module, as they can be routed
without having to be parsed completely.
The explicit registration of message handlers is something that I
thought would be tedious boilerplate code that could be automated away.
But experience now shows it to be the opposite. Allthough dull
repetition, the purpose it serves is to link together the code in a way
that humans can easily find. With explicit registration searching the
name of the message handler in the editor will show the registration that
ties it together with the type of message it handles and the kind of
registration. With implicit registration searching the name of the
message handler turns up only the handler with no clue as to what calls
it.
The role of a slave as a host for Factorio servers is not well reflected
in its name and has caused confusion among users. Rename slave to host
to better reflect its role as a host of Factorio servers.
The role of the master server as the party controlling every part of the
cluster is not well reflected in its name and has caused confusion among
users on several occasions. Rename master to controller to better
reflects its role.
Track summary statistic of total HTTP response durations using a
middleware that applies for all requests instead of having every request
needing to update the hit counter manually. Deprecates
endpointHitCounter.
Add /api/upload-save endpoint that streams the data via the slave
assigned to the target instance to the instance saves directory, along
with a ctl command and web interface button to do the upload.
Add authentication plugin that does a simple challenge response protocol
to verify that the player trying to log in is who they claim to be. The
process goes as follows:
1. A player code is generated for a player that wants to log in and is
displayed in-game.
2. The player code is input into the web interface and the master
generates a verification code as a challenge back along with a token.
3. The player inputs the verification code in-game and is logged in to
the web interface in doing so.
To do the actual login the web interface polls a verify endpoint every 2
seconds until the code has either been verified or expired. The
challenge comes from being unable to control the verification code and
having to input it as a player logging into one of the servers in order
to login.
Some additional steps are taken in order to increase the difficulty for
an attacker to subvert the challenge:
- The login process must be started from in-game.
- Generated codes are only valid for a limited timespan which is
configurable and 2 minutes by default.
- The verification code comes with a token that only works with one
player code and verification code combo and must be provided to pass
validation.
When reducing the collection interval in Prometheus it may be necessary
to to reduce the metrics gathering timeout to ensure it's less than the
interval. Add config options to adjust the timeouts used for metrics
gathering so that these can be set accordingly.
Add winston based logger for master, slave, ctl, lib, and web_ui code.
Logging done from slaves and instances are sent to the master server and
tagged with the instance and/or slave it originated from. This allows
for example error messages to propagate back to the source that iniated
the action. The master server also stores the log to allow it to be
queried at a later point. For slaves a copy of the slave's log is also
stored locally.
Add a sendRcon method to BaseInstancePlugin and the slave's Instance
class. This servers two purposes, one is to reduce the repetition when
send RCON commands particularly from plugins, and the other is to
instrument the RCON commands sent with metrics about duration and size.
Use lib* prefix for the modules impored from the lib package to help
distinguish them more from ordinary variables. Fix up the naming
convention for files to be consistently lowercase_underscore for
JavaScript files.
Make the config and link libraries compatible with packing for web with
webpack. This would fail due to ws not being web compatible and config
using a node function not available in the web shim for util.
Split up codebase into the packages @clsterio/lib, @clusterio/master,
@clusterio/slave, and @clusterio/ctl using lerna to manage them. This
will allow deployment to use only the parts needed reducing install
size, and better separate the code into shared library, master, slave
and control.
Turn the flat list of commands into a hierarchy of commands, converting
most required parameters into positional parameters and add plugin
loading support for clusterctl.
Implement notification to plugins for disconnect request, connect, drop,
and close. This allows plugins to properly handle slaves and the master
server shutting down without sending messages that will get lost.
Replace socket.io with plain WebSocket using the ws library as it
doesn't really provide any useful tools over plain WebSocket when it
comes to implementing reliable communication that works across
reconnects. The widespread browser support for WebSocket also means
it has little value to add for the web interface over WebSocket.
Daniel