exeunt
v1.1.2
Published
exiting a node.js process *and flushing stdout and stderr*
Downloads
871,814
Readme
node-exeunt
A module for (and discussion on) exiting a node.js process and flushing stdout and stderr.
Somewhere in the node.js 0.10 or 0.12 version range, and at least on certain
platforms including macOS and SmartOS, stdout and stderr stopped being
blocking. That means that where with node.js 0.10 or before your script might
write output and exit with process.exit([CODE])
, with newer versions of
node.js your output to stdout and/or stderr would sometimes not all get
written before the process exited. This is most commonly an annoyance for
command-line tools written in node.js, especially when used in a pipeline where
the problem more often manifests itself. The issue is surprisingly (at least to
me) complex. This repo will attempt to explain the tradeoffs with different
solutions and provide advice and one or more functions to use for exiting.
Usage
var exeunt = require('exeunt');
function main() {
// ...
exeunt(code); // flush stdout/stderr and exit
return; // `exeunt` returns, unlike `process.exit`
}
See the Solution 4 section below for details.
Note: exeunt()
is a small function. If you don't want yet another node
dependency, then feel free to just copy it to your repo.
The problem
A node.js script writes a lot of output (such that buffering occurs), and then exits. Not all output will be written before the process terminates. E.g.:
$ node examples/write-65k-and-exit.js | grep meta
[meta] start: writing 66560 bytes...
# 65k of output elided by the `grep`
[meta] done # all output was emitted this time
$ node examples/write-65k-and-exit.js | grep meta
[meta] start: writing 66560 bytes...
# the final 'done' line is missing
This example writes 65k to be more than the buffer size for a pipe (which is 64k, at least on macOS, IIUC). If we increase that to ~1MB, it is more frequent that output is truncated:
$ node examples/write-65k-and-exit.js 1000000 | grep meta
[meta] start: writing 1000000 bytes...
Solution 1: avoid process.exit
Summary: Use process.exitCode = code;
(added in node.js 0.12), do not use
process.exit([code])
, and ensure you have no active handles
(process._getActiveHandles()
).
Pros:
- All stdout and stderr content will be written before the node.js process exits. AFAIK this is the only solution that guarantees this.
Cons:
- You need to be diligent about closing active handles (from
setTimeout
,setInterval
, open sockets, etc.) otherwise your script will hang on exit. - In node 0.10 (if you need to support it), there is no way to exit with a
non-zero exit code without
process.exit(code)
.
Example showing an accidental hang on exit:
$ node examples/hang-because-active-handle.js | grep meta
[meta] start: writing 66560 bytes...
[meta] done
[meta] this interval is still running
[meta] this interval is still running
[meta] this interval is still running
^C
If you need to support node 0.10, here is a softExit()
function
that will use process.exitCode
if the node version supports it, else fallback
to process.exit
if necessary (with the potential for truncation).
Solution 2: give it a few seconds, then play hardball
Summary: Attempt to avoid process.exit, but set a timer to use it after a short while if it looks like we are hanging.
Pros:
- In correct operation, your script will write out all stdout/stderr before exiting.
Cons:
- If stdout/stderr takes more than 2s (or whatever timeout you choose) to flush, then output will still be truncated. This is the main tradeoff to avoid a hang.
- This technique involves calling a function that doesn't synchronously
exit the process like
process.exit()
does. That means you need to handle it returning and code still executing. That might be as simple as callingreturn;
, or it might be more difficult. It depends on your application's code.
$ node examples/hardball-after-2s.js | grep meta
[meta] start: writing 66560 bytes...
[meta] done
[meta] this interval is still running
[meta] this interval is still running
[meta] hardball exit, you had your chance
Solution 3: set stdout/stderr to be blocking
This all started because stdout/stderr weren't blocking. Let's just set them to be blocking again.
Pros:
- Stdout and stderr will be flushed as soon as your script writes to them.
Cons:
- The node event loop can block if the other end of those pipes isn't reading! This was a subtlety that surprised me. See https://gist.github.com/misterdjules/3aa4c77d8f881ffccba3b6e6f0050d03 for an example showing this. (TODO: include those scripts in examples/ here.)
$ node examples/set-blocking-write-65k-and-exit.js 1000000 | grep meta
[meta] start: writing 1000000 bytes...
[meta] done
Solution 4: exeunt
Set stdout/stderr to be blocking, but only when about to exit.
Usage:
var exeunt = require('exeunt');
function main() {
// ...
exeunt(code); // flush stdout/stderr and exit
return; // `exeunt` returns, unlike `process.exit`
}
Pros:
- Stdout and stderr will most likely (see below) be flushed before exiting.
- Because
exeunt()
is callingprocess.exit()
, there is no special issue with the node event loop blocking.
Cons:
exeunt()
callsprocess.exit()
asynchronously (insetImmediate
), which means you need to handle code still executing. Depending on how your code is structured, that might just require callingreturn;
.process.exit
is called insetImmediate
to ensure that one more pass through the event loop will flush stdout/stderr. That event loop pass will also run timers (as part ofuv__run_timers()
inuv_run()
). I.e. currentsetTimeout
s andsetIntervals
may run one more time. My expectation is that this shouldn't be a practical concern for most programs, but it might be for yours.
$ node examples/write-65k-and-exeunt.js 1000000 | grep meta
[meta] start: writing 1000000 bytes...
[meta] done
The code, to show what is happening, is here: https://github.com/joyent/node-exeunt/blob/master/lib/exeunt.js#L59-L87. There are some subtleties.
First, we can't just exit synchronously:
setBlocking();
process.exit(code);
because that will synchronously call the exit syscall, and the process will
terminate, before any IO handling to write buffered stdout/stderr. Instead
we use setImmediate
to ensure that there is one more run through the
node event loop which calls
uv__io_poll
to service IO requests before calling our setImmediate
handler.
Second, we said that stdout/stderr will "most likely be flushed" above, because
it appears that uv__io_poll
is
tuned
to limit the amount of events if will handle in a single event loop pass. I
haven't yet come up with example code that hits this threshold, however.
Open Questions
We haven't verified all our observations yet. This section includes Rumsfeldian known unknowns.
We need to verify the observations I've made above. At time of writing I was testing out the above examples with node v4.8.0 on macOS 10.11.6.
What are the conditions in libuv's
uv__io_poll
(which is called once for each pass through the node event loop) such that thecount = 48
guard is triggered, such that not all IO is handled in that last pass? https://github.com/nodejs/node/blob/v4.8.0/deps/uv/src/unix/kqueue.c#L291 https://github.com/nodejs/node/blob/v4.8.0/deps/uv/src/unix/sunos.c#L287 Understanding this would be useful to know if and what limitation there is on solution 4.Test yargs' cases using setBlocking, e.g. https://github.com/yargs/yargs/blob/8756a3c63dfd2ceae303067b46075de5c982af66/yargs.js#L1010-L1012 to see if they work.
See Also
nodejs/node#6980 "Tracking issue: stdio problems". The node.js core issue that aims to be the tracker for issues related to this. Aside: One of the linked issues includes this:
If this is currently breaking your program, please use this temporary fix:
[process.stdout, process.stderr].forEach((s) => { s && s.isTTY && s._handle && s._handle.setBlocking && s._handle.setBlocking(true) })
I believe the
s.isTTY
guard needs to be dropped.nodejs/node@ab3306a is the commit where a TTY is set to blocking. This is why (at least for node releases with this commit), stdout/stderr flushing is not an issue for a node app called interactively and without piping into another program.
https://github.com/yargs/set-blocking is a small module related to the same problem. It states: "In yargs we only call setBlocking(true) once we already know we are about to call process.exit(code)." This is therefore similar to "Solution 4" described here, and the provided
exeunt()
function. It isn't clear to me all of yargs' usages of this pattern callprocess.exit
in a separate tick, which is necessary to actually flush output.
License
MPL 2.0