Skip to content

Latest commit

 

History

History
788 lines (593 loc) · 16.4 KB

presentation.md

File metadata and controls

788 lines (593 loc) · 16.4 KB
Raw

autoscale: true build-lists: true

[fit] Robust

[fit] Error Handling

[fit] in Node.js

Who are you?

  • I'm Lewis, I like JavaScript (and other things)
  • Made Bee-Queue, Redis-backed job queue for Node.js
    • Like Celery, Resque, Kue, Bull
  • Worked on raven-node, Sentry's Node error reporting SDK
    • It captures and reports about 100 million errors per week

Why should you want robust error handling in Node?

This talk is for you if...

You've ever written code like this:

fs.readFile('myFile.txt', { encoding: 'utf8' }, function (err, data) {
  console.log(data);
});

^ if you've ever written node code, you've written node code like this

Or this:

http.get(myUrl, function (res) {
  doSomethingWith(res);
});

Or even this:

try {
  data = JSON.parse(userInput);
} catch (e) {
  // this should never happen
}
doSomethingWith(data);

"this should never happen"

Famous last words

We can do better!

This is a little better:

fs.readFile('myFile.txt', { encoding: 'utf8' }, function (err, data) {
  if (err) return console.error(err);
  doSomethingWith(data);
});

But we can do a lot better

Overview of topics

  • Background and how Node is special/different
  • Handle what you can, avoid what you can't
  • The robust game plan to follow
  • Exceptions, Callbacks, Promises, EventEmitters, and more
  • Catching, reporting, shutting down, restarting gracefully

Error Mechanisms in Other Languages

  • Python: try/except and raise
  • Ruby: begin/rescue and raise
  • PHP: try/catch and throw
  • Lua: pcall() and error()

So Node has some similar thing, right?

Well, sure, this works:

try {
  throw new Error('boom!');
} catch (e) {
  console.log('Aha! I caught the error!');
}

Well, sure, this works:

try {
  throw new Error('boom!');
} catch (e) {
  console.log('Aha! I caught the error!');
}
$ node try-catch.js

Well, sure, this works:

try {
  throw new Error('boom!');
} catch (e) {
  console.log('Aha! I caught the error!');
}
$ node try-catch.js
> Aha! I caught the error!

But where try-catch comes up short:

try {
  setTimeout(function () {
    throw new Error('boom!');
  }, 0);
} catch (e) {
  console.log('Aha! I caught the error!');
}

But where try-catch comes up short:

try {
  setTimeout(function () {
    throw new Error('boom!');
  }, 0);
} catch (e) {
  console.log('Aha! I caught the error!');
}
$ node try-settimeout.js

But where try-catch comes up short:

try {
  setTimeout(function () {
    throw new Error('boom!');
  }, 0);
} catch (e) {
  console.log('Aha! I caught the error!');
}
$ node try-settimeout.js
/Users/lewis/dev/node-error-talk/try-settimeout.js:3
    throw new Error('boom!');
    ^
Error: boom!
    at Timeout._onTimeout (/Users/lewis/dev/node-error-talk/try-settimeout.js:3:11)
    at ontimeout (timers.js:365:14)
    at tryOnTimeout (timers.js:237:5)
    at Timer.listOnTimeout (timers.js:207:5)

Why?

  • Try-catch is synchronous, setTimeout is asynchronous
  • Callback is queued and will throw the error later
  • Catch block is no longer waiting to catch the exception
  • Run-to-completion semantics & event loop are behind this

How Node is Different

  • Other languages:
    • Each process handles one request at a time
    • Everything is synchronous, try/catch works fine
    • Easy to keep one request from blowing everything up
  • Node: single process, cooperative concurrency, asynchronous I/O
    • Handles multiple requests at the same time in the same thread
    • Try/catch doesn't work well in asynchronous world
    • Any one request blowing up could mess up the others

What this means for us

  • We need other mechanisms to handle asynchronous errors
  • We can't try/catch and throw exceptions for everything
  • An error in one request can take down the entire server
  • We have to be extra careful to keep things online
  • Our program can end up in unknown states
    • Only correct thing to do might be shut down!

^ point out use of errors vs exceptions

Errors vs Exceptions

What's the difference, anyway?

Errors

Error is just a special class in JavaScript

  • You can pass an Error object around like any other value
  • Runtime errors throw an Error object
  • Has a message and a stack property
  • Also RangeError, ReferenceError, SyntaxError, others
var myError = new Error('my error message');

^ Mention engines have special treatment: V8 etc will populate stack trace based on line where Error is instantiated

Exceptions

Exception: what happens when you throw something

  • Usually you throw an Error object
  • Call stack unwinds looking for a catch block
  • You can throw anything, not just Error objects...but don't
  • If an exception is unhandled, Node will shut down
throw new Error('something bad happened!');
throw 'something bad happened' // avoid this

^ Mention more on unhandled later?

Stack trace example

function a() {
  // call stack here is [a]
  b();
}

function b(x) {
  // call stack here is [b, a]
  try {
    c()
  } catch (e) {
    console.log(e.stack);
  }
}

function c() {
  // call stack here is [c, b, a]
  throw new Error('boom');
}

a();
Error: boom
    at c (/Users/lewis/dev/node-error-talk/try-catch.js:14:9)
    at b (/Users/lewis/dev/node-error-talk/try-catch.js:7:5)
    at a (/Users/lewis/dev/node-error-talk/try-catch.js:2:3)
    at Object.<anonymous> (/Users/lewis/dev/node-error-talk/try-catch.js:17:1)
    ...
    <more node core module frames>

This is useful! We want to see it!

Errors vs Exceptions in Async Node Land

  • We're generally not going to throw Exceptions
  • We're going to pass Error objects around a lot

Three Guiding Principles

Always know when your errors happen

Avoid patterns like:

function (err, result) {
  if (err) { /* drat, ignore */ }
}

try { ... } catch (e) {
  // this should never happen
}

Promise.catch(function (reason) {
  // surely this won't happen
});

req.on('error', function (err) {
  // oh well, not gonna do anything
});

Handle what you can, avoid what you can't

Operational Errors vs Programming Errors

  • Operational error: recoverable - expect and handle these
    • Typically an Error object being passed around
  • Programming error: nonrecoverable - try to avoid these
    • Typically an Exception thrown

^ There's no amount of additional code you can write to recover from a typo

Operational Errors to Expect

  • Network timeouts
  • Database is down
  • Disk got full
  • 3rd party API returning errors: S3 goes down
  • Unexpected or missing user inputs (JSON.parse)

^ Tubes get jammed ^ S3 goes down ^ Users do things you don't expect

Programming Errors to Avoid

  • Silently ignoring/swallowing errors instead of handling them
  • Classic JavaScript errors
    • TypeError: undefined is not a function
    • TypeError: Cannot read property 'x' of undefined
    • ReferenceError: x is not defined
  • Invoking a callback twice
  • Using the wrong error mechanism in the wrong place

^ Linters can literally forbid you from ignoring errors ^ Any time you use a . you should feel uncomfortable: is it defined? ^ Don't throw when you should callback

What to do with each:

  • Operational errors

    • Known; handle manually wherever they may occur
    • Recoverable if handled correctly: S3 being down doesn't kill us
    • Avoid assuming anything is reliable outside your own process

  • Programming errors

    • Unknown; catch with global error handler
    • Nonrecoverable: we're gonna have to abandon ship
      • No amount of additional code can fix a typo
    • Use a linter to help avoid many common problems

^ Operational errors that we don't handle become programming errors ^ Highlight "unknown" bit about programming errors ^ No amount of additional code can fix a typo

Don't keep running in an unknown state

Don't keep running in an unknown state

  • State shared across multiple requests: less isolation
  • Unexpected error in one request can pollute state of others
  • Polluted state can lead to undefined behavior
    • Memory leaks, infinite loops, security issues
  • Only way to get back to a known good state: bail out, restart

^ Understanding these guidelines based on what's unique about node leads us to...

This leads us to the game plan

The game plan

  1. Follow the guiding principles
  2. Know and use different mechanisms for effective handling
  3. Have a global catch-all for the errors you couldn't handle
  4. Use a process manager so shutting down is no big deal
  5. Accept when it's time to pack up shop, clean up, shut down

^ In accordance with our three guiding principles... ^ - Always know when errors happen: don't ignore! ^ - Handle what you can, avoid what you can't ^ - Don't keep running in an unknown state

1. Know and use different mechanisms for effective handling

Error Mechanisms in Node

  • Try/Catch - throw and try/catch
  • Callbacks - err first argument and if (err)
  • Promises - reject(err) and .catch()
  • Async/Await - Sugar for promises + try/catch
  • EventEmitters - error events and .on('error')
  • Express - next(err) and error-handling middleware

^ Touched on first two, probably seen them before

Callbacks & Try-catch

function readAndParse(file, callback) {
  fs.readFile(file, { encoding: 'utf8' }, function (err, data) {
    if (err) return callback(err);
    var parsed;
    try {
      parsed = JSON.parse(data);
    } catch (e) {
      return callback(e);
    }
    callback(null, parsed);
  });
}

Promises

var p = new Promise(function (resolve, reject) {
  fs.readFile('data.txt', function (err, data) {
    if (err) return reject(err);
    resolve(data);
  });
});
p.then(parseJson)
.then(doSomethingElse)
.catch(function (reason) {
  // if readFile or JSON parsing or something else failed,
  // we can handle it here
});

^ reason should be an Error object but promise terminology is "reason" not "err"

Async/Await

try {
  await somePromiseThatRejects()
} catch (e) {
  // e is the rejection reason!
}

EventEmitters

  • Servers, sockets, requests, streams
    • Long-lived objects with asynchronous stuff going on
  • They can emit error events: listen for them!
  • If an error event is emitted without an error listener...
    • The Error object will be thrown instead!
    • How operational errors become programming errors

EventEmitter Request Example

var req = http.get(url, function (res) {
  doSomething(res);
});

req.on('error', function (err) {
  // we caught the request error, let's recover
}

Express Error Middleware

app.post('/login', function (req, res, next) {
  db.query('SELECT ...', function (err, user) {
    if (err) return next(err);
  });
});

app.use(function (req, res, next, err) {
  // spit out your own error page, log the error, etc
  next();
});

^ Make sure to know how your libraries and dependencies will communicate errors so you can handle operational errors!

2. Have a global catch-all for the errors you couldn't handle

Basic Global Error Handler

process.on('uncaughtException', function (err) {
  console.log('Uncaught exception! Oh no!');
  console.error(err);
  process.exit(1);
});

^ Remember: we have to know the error happened! ^ Don't try to recover at this point: it's too late

3. Use a process manager so shutting down is no big deal

Process managers and Node

  • Run multiple server processes
    • One of them dying won't take us offline
    • Node cluster module
  • Process managers: systemd, pm2, forever, naught
    • Will automatically restart processes when they die
    • Some provide further Node-specific functionality

Example with naught

var server = http.createServer(...);

process.on('uncaughtException', function (err) {
  console.log('Uncaught exception! Oh no!');
  console.error(err);
  // tell naught to stop sending us connections & start up a replacement
  process.send('offline');
  process.exit(1);
});

server.listen(80, function () {
  // tell naught we're ready for traffic
  if (process.send) process.send('online');
});

^ more later on things shutDownGracefully might do

4. Accept when it's time to pack up shop, clean up, shut down

When we catch a "fatal" error, we want to:

  • Quit accepting new connections
  • Start reporting whatever we're gonna report
  • Tell proc manager we're gonna die so it starts replacement
  • Wait for any existing requests, sockets, etc to be dealt with
  • Close any open resources, connections, etc
  • Shut down

We want to report that stack trace!

function reportError(err, cb) {
  // send the stack trace somewhere, then call cb()
}

process.on('uncaughtException', function (err) {
  process.send('offline');
  reportError(err, function (sendErr) {
    // once error has been reported, let's shut down
    process.exit(1);
  });
});

Maybe get a text message:

var myPhone = "..."
function reportError(err, cb) {
  console.error(err);
  twilio.sendTextMessage(myPhone, err.message, cb);
}

Also useful to report operational errors

db.query('SELECT ...', function (err, results) {
  if (err) {
    return reportError(err);
  }
  doSomething(results);
});

But then the database goes down...

Alternatively, use Sentry

The raven npm package is Sentry's Node error reporting SDK:

var Raven = require('raven');
Raven.config('<my-sentry-key>');

function reportError(err, cb) {
  console.error(err);
  Raven.captureException(err, cb);
}

Graceful shutdown

server = http.createServer(...);
function shutDownGracefully(err, cb) {
  // quit accepting connections, clean up any other resources
  server.close(function () {
    // could also wait for all connections: server._connections
    reportError(err, cb)
  });
}

process.on('uncaughtException', function (err) {
  process.send('offline');
  shutDownGracefully(function () {
    process.exit(1);
  });
});
  • Wait for any existing requests, sockets, etc to be dealt with
  • Close any open resources, connections, etc

Big Combined Example

server = http.createServer(...);

function reportError(err, cb) {
  console.error(err);
  Raven.captureException(err, cb);
}

function shutDownGracefully(err, cb) {
  // quit accepting connections, clean up any other resources
  server.close(function () {
    // could also wait for all connections: server._connections
    reportError(err, cb)
  });
}

process.on('uncaughtException', function (err) {
  process.send('offline');
  shutDownGracefully(function () {
    process.exit(1);
  });
});

server.listen(80, function () {
  if (process.send) process.send('online');
});

What NOT to do with a global catch-all:

  • Just log the error and carry on
  • Keep the process running indefinitely
  • Try to recover in any way: it's too late!
  • Try to centralize handling of operational errors into one place

Other global error mechanisms

  • process.on('uncaughtException')
  • process.on('unhandledRejection')
    • Currently non-fatal, warning starting in Node 7
    • Future: fatal, will cause process exit
  • Domains: application code shouldn't need them

Recap: overall

  1. Follow the guiding principles
  2. Know and use different mechanisms for effective handling
  3. Have a global catch-all for the errors you couldn't handle
  4. Use a process manager so shutting down is no big deal
  5. Accept when it's time to pack up shop, clean up, shut down

Related things I didn't go into

  • Run-to-completion semantics & the event loop
  • V8 stacktrace API
  • The "callback contract"
  • Asynchronous stacktraces
  • Cluster module & process managers
  • Domains
  • async_hooks (!!!!!)

^ Each of these bullets is at least a lightning talk worth of stuff

Some related links

Thank you!