SQLite has a worse-is-better philosophy about automatically converting
between different data types, such as strings and floating point numbers:
https://www.sqlite.org/quirks.html#flexible_typing
This means querying for the string "1.10" in a given column can return
rows where the column is actually the string "1.1", since SQLite imagines
you might be talking about the number 1.1, rather than the string you
actually requested.
This "feature" became a problem for Meteor after we published Meteor 1.10,
which caused SQLite to return multiple rows for the getReleaseVersion
query, including both Meteor 1.10 and Meteor 1.1 (which is ancient, from
March 2015).
While this behavior seems completely indefensible, the SQLite
documentation clearly does not consider it a bug, which forces us to work
around the consequences by double-checking the queried results with the
filterExactRows helper function.
This fixes the bug where commands like `meteor add-platform ios` would
fail the first time with an error that cordova-lib could not be found,
even though we attempt to install the necessary packages if they have not
already been installed.
To make a very long story short, calling moduleDoesResolve before
installing dependencies like cordova-lib was causing Node.js to cache the
_absence_ of cordova-lib/package.json permanently in the new
packageJsonCache, which cannot be invalidated or cleared by user code:
f8f20892e9/lib/internal/modules/cjs/loader.js (L245-L255)
Although we could potentially propose a change to Node to allow the
packageJsonCache to be invalidated, a more immediate solution is simply to
avoid calling moduleDoesResolve when there's any chance the module will
not resolve. Because we still want to avoid repeatedly installing Cordova
dependencies every time we run a Cordova command, we instead check whether
the necessary dependencies are installed by examining the file system.
I considered using a nonzero process.exit code, but I didn't want to run
any risk of reusing a meaningful code, and the accepted range of codes
unfortunately does not include parseInt("reload", 36), or 1657112629.
Should fix#10934.
These stray EPIPE messages have been a problem for a while now, especially
on Windows.
In most cases, they appear to stem from sending a message to a process
that is about to exit, such that the message races against the exit, and
sometimes loses. When that message is a just an obligatory response to a
final message from the exiting process, the exiting process probably does
not care about (and will not receive) the response, so we can safely
swallow the EPIPE error.
Now that all otherProcess.send calls have a callback function, I'm hopeful
we will never see EPIPE errors again.
