This hack dates all the way back to 2013: a2c4a78743
Though it is convenient to reload the browser when server files change
while running test-packages, that's not the behavior of most Meteor apps
that use the autoupdate package, and this hack introduced a signficant
difference in behavior between the test-in-browser and test-in-console
driver packages, which finally surfaced due to the interaction between
@toinevk's headless testing PR #9814 and my refactoring of the autoupdate
package (fe9e4035f9). Tests should behave
the same regardless of which driver package is used.
It turns out there's a better way to make the browser reload each time the
server restarts: simply modify Meteor.settings.public, since that object
is included in the client hashes computed by the webapp package.
It seems obvious in hindsight, but any logic relating to the
AUTOUPDATE_VERSION override should reside within the autoupdate package,
and the true client hashes should be available to any other package that
needs them, without AUTOUPDATE_VERSION getting in the way.
Every process is potentially the child of some other process and the
parent of zero or more child processes of its own, so it's confusing to
use terminology that always treats the current global.process as a
"parent" process, or to include PARENT and CHILD in the message types.
Instead, this new implementation uses message types MESSAGE, RESPONSE,
PING, and PONG, and refers to `process` and `otherProcess` objects,
with the caveat that sometimes `process === otherProcess`, because
`process.send` can be used to send messages to the parent process.
Instead of relying on the child to send a special CHILD_READY message to
the parent when it's ready to receive messages, the sending process polls
the receiving process with a preflight PING message, and the receiving
process immediately responds with a PONG when ready.
Fixes#10073, per
https://github.com/meteor/meteor/issues/10073#issuecomment-405290391
While thinking about this bug, I realized that sending IPC messages to
specific packages in the server process was much less flexible than
sending messages based on an arbitrary topic string, since the topic
string approach allows both `autoupdate` and `dynamic-import` to listen
for the same message.
The topic string approach calls for a listener interface like
`onMessage(topic, callback)`, which elegantly replaces the previous
approach of requiring packages to export a single `onMessage` function.
However, because the `meteor` package does not have access to the module
system, implementing the `onMessage` listener interface in the `meteor`
package would have required exposing an API like `Meteor.onMessage(topic,
callback)`, which has an unpleasant global smell to it. Instead, the
`onMessage` function should be explicitly imported (using the module
system) from a less-generically-named package.
Since I knew I was going to have to move the message dispatch logic out of
the `meteor` package, I decided to create a new package called
`inter-process-messaging` to implement the parent/child components of the
IPC system.
https://github.com/meteor/meteor/pull/10055#discussion_r201855997
As I explained in this comment, Package._on(packageName, callback) was a
bad API because it never called the callback if the package was not
installed, which caused any app not using the autoupdate package to get
stuck trying to communicate with the autoupdate package.
Instead of having every message consumer listen to every message and act
on the ones that seem relevant to its interests, we now have a single
process.on("message", callback) hook that can dispatch messages to
different Meteor packages running in the server process.
Receiving packages should export an onMessage function. The onMessage
function may be async, and its result will be delivered back to the build
process as the result of the sendMessage Promise.
Now that we're postponing the legacy build until after the first client
refresh message is sent, there's a risk that changes to the legacy build
will not be picked up until after the next rebuild.
If we attempted to fix that problem by sending the refresh message after
the legacy bundle is rebuilt, then we would lose most of the benefit of
delaying the legacy build, because the client would not refresh until
after the legacy build completed.
The right way to fix the problem is by sending a second client refresh
message after the legacy build finishes, but doing so with the current
autoupdate implementation would very likely cause modern clients to reload
a second time.
The solution implemented by this commit is simple in theory: the
autoupdate package should keep track of distinct versions for each client
architecture, so that modern clients will refresh only when the modern
versions change, and legacy clients will refresh only when the legacy
versions change, which allows us to send two refresh messages without
causing any clients to refresh more than once.
In reality, this was a fairly major rewrite, since the ClientVersions
collection has a totally different schema now. I've tested it as well as I
can, though I'm not entirely sure what will happen if clients using the
previous version of the autoupdate package begin receiving DDP messages
from this version of the autoupdate server code.
This means the autoupdate package is no longer responsible for recomputing
client hashes, and we can recompute the hashes whenever there's a new (or
updated) client program, which enables delayed builds of architectures
like web.browser.legacy (#10055).
