Hello,
I use Node.js to provide an API for storing data on a MongoDB database.
I ran multiple tests on a read method, which takes ids and returns the corresponding documents. The point is that I must return these documents in the specified order. To ensure that, I use the following code:
// Sequentially fetch every element
function read(ids, callback) {
var i = 0;
var results = [];
function next() {
db.findOne(ids[i], function (err, doc) {
results.push(err ? null : doc);
if (ids.length > ++i) {
return next();
}
callback(results);
});
}
next();
}
This way, documents are fetched one-by-one, in the right order. It takes about 11s on my laptop to retrieve 27k documents.
However, I thought that it was possible to improve this method:
// Asynchronously map the whole array
var async = require('async');
function read(ids, callback) {
async.map(ids, db.findOne.bind(db), callback):
}
After running a single test, I was quite satisfied seeing that the 27k documents were retrieved in only 8s using simpler code.
The problem happens when I repeat the same request: the response time keeps growing (proportionally to the number of elements retrieved): 9s 10s 11s 12s.... This problem does not happen in the sequential version.
I tried two versions of Node.js, v6.2.0 and v0.10.29. The problem is the same. What causes this latency and how could I suppress it?
Try to use async.mapLimit to prevent overload. You need some tests to tune limit value with your environment.
But find({_id: {$in: list}}) is always better, because single database request instead of multiple.
I suggest you to try to perform restore of original order client-side.
Something like this:
function read(ids, cb) {
db.find(
{_id: {$in: ids.map(id => mongoose.Types.ObjectId(id))}},
process
);
function process(err, docs) {
if (err) return cb(err);
return cb(null, docs.sort(ordering))
}
function ordering(a, b) {
return ids.indexOf(b._id.toString()) - ids.indexOf(a._id.toString());
}
}
May be, find query needs to be corrected, I can't to know what exact mongodb driver you use.
This code is first-try, more manual sorting can improve performance alot. [].indexOf is heavy too(O(n)).
But I'm almost sure, even as-is now, it will work much faster.
Possible ordering replacement:
var idHash = {};
for(var i = 0; i < ids.length; i++)
idHash[ids[i]] = i;
function ordering(a, b) {
return idHash[b._id.toString()] - idHash[a._id.toString()];
}
Any sort algorithm has O(nlogn) in best case, but we already know result position of each found document, so, we can restore original order by O(n):
var idHash = ids.reduce((c, id, i) => (c[id] = i, c), {});
function process(err, docs) {
if (err) return cb(err);
return cb(null,
docs.reduce(
(c, doc) => (c[idHash[doc._id.toString()]] = doc, c),
ids.map(id => null))) //fill not_found docs by null
}
Functional style makes code flexier. For example this code can be easy modified to use async.reduce to be less sync-blocking.
I have an orientdb database. I want to use nodejs with RESTfull calls to create a large number of records. I need to get the #rid of each for some later processing.
My psuedo code is:
for each record
write.to.db(record)
when the async of write.to.db() finishes
process based on #rid
carryon()
I have landed in serious callback hell from this. The version that was closest used a tail recursion in the .then function to write the next record to the db. However, I couldn't carry on with the rest of the processing.
A final constraint is that I am behind a corporate proxy and cannot use any other packages without going through the network administrator, so using the native nodejs packages is essential.
Any suggestions?
With a completion callback, the general design pattern for this type of problem makes use of a local function for doing each write:
var records = ....; // array of records to write
var index = 0;
function writeNext(r) {
write.to.db(r, function(err) {
if (err) {
// error handling
} else {
++index;
if (index < records.length) {
writeOne(records[index]);
}
}
});
}
writeNext(records[0]);
The key here is that you can't use synchronous iterators like .forEach() because they won't iterate one at a time and wait for completion. Instead, you do your own iteration.
If your write function returns a promise, you can use the .reduce() pattern that is common for iterating an array.
var records = ...; // some array of records to write
records.reduce(function(p, r) {
return p.then(function() {
return write.to.db(r);
});
}, Promsise.resolve()).then(function() {
// all done here
}, function(err) {
// error here
});
This solution chains promises together, waiting for each one to resolve before executing the next save.
It's kinda hard to tell which function would be best for your scenario w/o more detail, but I almost always use asyncjs for this kind of thing.
From what you say, one way to do it would be with async.map:
var recordsToCreate = [...];
function functionThatCallsTheApi(record, cb){
// do the api call, then call cb(null, rid)
}
async.map(recordsToCreate, functionThatCallsTheApi, function(err, results){
// here, err will be if anything failed in any function
// results will be an array of the rids
});
You can also check out other ones to enable throttling, which is probablya good idea.
While coding in Node.js, I encountered many situations when it is so hard to implement some elaborated logic mixed with database queries (I/O).
Consider an example written in python. We need to iterate over an array of values, for each value we query the database, then, based on the results, we need to compute the average.
def foo:
a = [1, 2, 3, 4, 5]
result = 0
for i in a:
record = find_from_db(i) # I/O operation
if not record:
raise Error('No record exist for %d' % i)
result += record.value
return result / len(a)
The same task in Node.js
function foo(callback) {
var a = [1, 2, 3, 4, 5];
var result = 0;
var itemProcessed = 0;
var error;
function final() {
if (itemProcessed == a.length) {
if (error) {
callback(error);
} else {
callback(null, result / a.length);
}
}
}
a.forEach(function(i) {
// I/O operation
findFromDb(function(err, record) {
itemProcessed++;
if (err) {
error = err;
} else if (!record) {
error = 'No record exist for ' + i;
} else {
result += record.value;
}
final();
});
});
}
You can see that such code much harder to write/read, and it is more prone to errors.
My questions:
Is there a way to make above Node.js code cleaner?
Imagine more sophisticated logic. For example, when we obtained a record from the db, we might need do another db query based on some conditions. In Node.js that becomes a nightmare. What are common patterns for dealing with such tasks?
Based on your experience, does the performance gain deserves the productivity loss when you code with Node.js?
Is there other asynchronous I/O framework/language that is easier to work with?
To answer your questions:
There are libraries such as async which provide a variety of solutions for common scenarios when working with asynchronous tasks. For "callback hell" concerns, there are many ways to avoid that as well, including (but not limited to) naming your functions and pulling them out, modularizing your code, and using promises.
More or less what you currently have is a fairly common pattern: having counter and function index variables with an array of functions to call. Again, async can help here because it reduces this kind of boilerplate that you will probably find yourself repeating often. async currently doesn't have methods that really allow for skipping individual tasks, but you could easily do this yourself if you are writing the boilerplate (just increment the function index variable by 2 for example).
From my own experience, if you properly design your javascript code with asynchronous in mind and use a lot of tools like async, you will find it easier to develop with node. Writing for asynchronous vs synchronous in node is typically always going to be more complicated (although less so with generators, fibers, etc. as compared to callbacks/promises).
I personally think that deciding on a language based upon that single aspect is not worthwhile. You have to consider much much more than just the design of the language, for example the size of the community, availability of third party libraries, performance, technical support options, ease of code debugging, etc.
Just write your code more compactly:
// parallel version
function foo (cb) {
var items = [ 1, 2, 3, 4, 5 ];
var pending = items.length;
var result = 0;
items.forEach(function (item) {
findFromDb(item, function (err, record) {
if (err) return cb(err);
if (!record) return cb(new Error('No record for: ' + item))
result += record.value / items.length;
if (-- pending === 0) cb(null, result);
});
});
}
That clocks in at 13 source lines of code compared to the 9 sloc for python that you posted. However, unlike the python that you posted, this code runs all the jobs in parallel.
To do the same thing in series, a trick I usually do is a next() function defined inline that invokes itself and pops a job off of an array:
// sequential version
function foo (cb) {
var items = [ 1, 2, 3, 4, 5 ];
var len = items.length;
var result = 0;
(function next () {
if (items.length === 0) return cb(null, result);
var item = items.shift();
findFromDb(item, function (err, record) {
if (err) return cb(err);
if (!record) return cb(new Error('No record for: ' + item))
result += record.value / len;
next();
});
})();
}
This time, 15 lines. The nice thing is that you can easily control whether the actions should happen in parallel or sequentially or somewhere in between. That is not so easy in a language like python where everything is synchronous and you've got to do lots of work-arounds like threads or evented libraries to get things back up to asynchronous. Try implementing a parallel version of what you have in python! It would most certainly be longer than the node version.
As for the promise/async route: it's not actually all that hard or bad to use ordinary functions for these relatively simple kinds of tasks. In the future (or in node 0.11+ with --harmony) you can use generators and a library like co, but that feature isn't widely deployed yet.
Everyone here seems to be suggesting async, which is a great library. But to give another suggestion, you should take a look at Promises , which is a new built-in being introduced to the language (and currently has several very good polyfills). It allows you to write asynchronous code in a way that looks much more structured. For example, take a look at this code:
var items = [ 1, 2, 3, 4 ];
var processItem = function(item, callback) {
// do something async ...
};
var values = [ ];
items.forEach(function(item) {
processItem(item, function(err, value) {
if (err) {
// something went wrong
}
values.push(value);
// all of the items have been processed, move on
if (values.length === items.length) {
doSomethingWithValues(values, function(err) {
if (err) {
// something went wrong
}
// and we're done
});
}
});
});
function doSomethingWithValues(values, callback) {
// do something async ...
}
Using promises, it would be written something like this:
var items = [ 1, 2, 3, 4 ];
var processItem = function(item) {
return new Promise(function(resolve, reject) {
// do something async ...
});
};
var doSomethingWithValues = function(values) {
return new Promise(function(resolve, reject) {
// do something async ...
});
};
// promise.all returns a new promise that will resolve when all of the promises passed to it have resolved
Promise.all(items.map(processItem))
.then(doSomethingWithValues)
.then(function() {
// and we're done
})
.catch(function(err) {
// something went wrong
});
The second version is much cleaner and simpler, and that barely even scratches the surface of promises real power. And, like I said, Promises are in es6 as a new language built-in, so (eventually) you won't even need to load in a library, it will just be available.
don't use anonymous (un-named) functions they make the code ugly and they make debugging much harder, so always name your functions and define them outside the function scope not inline.
that is a real issue with Node.js (it is called callback hell or pyramid of doom ,..) you can solve this issue by using promises or using async.js which have so many functions for handling different situations (waterfall, parallel, series, auto, ...)
well the performance gain is absolutely a good thing and it is not that much loss (when you start to master it) and also the Node.js community is great.
Check async.js, q.
The more I work with async the more I love it and I like node more. Let me give you a simple example of what I have for a server initialization.
async.parallel ({
"job1": loadFromCollection1,
"job2": loadFromCollection2,
},
function (initError, results) {
if (initError) {
console.log ("[INIT] Server initialization error occurred: " + JSON.stringify(initError, null, 3));
return callback (initError);
}
// Do more stuff with the results
});
In fact, this very same approach can be followed and one can pass different arguments to the different functions that correspond to the various jobs; see for example Passing arguments to async.parallel in node.js.
To be perfectly honest with you, I prefer the node-way which is also non-blocking. I think node forces someone to have a better design and sometimes you spend time creating more definitions and grouping functions and objects in arrays so that you can write better code. The reason I think is that in the end you want to exploit some variant of async and mix and merge stuff accordingly. In my opinion, spending some extra time and thinking about the code a bit more is well worth it when you also take into account that node is asynchronous.
Other than that, I think it is a habit. The more one writes code for node, the more one improves and writes better asynchronous code. What is good on node is that it really forces someone to write more robust code since one starts respecting all the error codes from all the functions much more. For example, how often do people check, say if malloc or new have succeeded and one does not have an error handler for a NULL pointer after the command has been issued? Writing asynchronous code though forces one to respect the events and the error codes that the events have. I guess one obvious reason is that one respects the code that one writes and in the end we have to write code that returns errors so that caller knows what happened.
I really think that you need to give it more time and start working with async more. That's all.
"If you try to code bussiness db login using pure node.js, you go straight to callback hell"
I've recently created a simple abstraction named WaitFor to call async functions in sync mode (based on Fibers): https://github.com/luciotato/waitfor
check the database example:
Database example (pseudocode)
pure node.js (mild callback hell):
var db = require("some-db-abstraction");
function handleWithdrawal(req,res){
try {
var amount=req.param("amount");
db.select("* from sessions where session_id=?",req.param("session_id"),function(err,sessiondata) {
if (err) throw err;
db.select("* from accounts where user_id=?",sessiondata.user_ID),function(err,accountdata) {
if (err) throw err;
if (accountdata.balance < amount) throw new Error('insufficient funds');
db.execute("withdrawal(?,?),accountdata.ID,req.param("amount"), function(err,data) {
if (err) throw err;
res.write("withdrawal OK, amount: "+ req.param("amount"));
db.select("balance from accounts where account_id=?", accountdata.ID,function(err,balance) {
if (err) throw err;
res.end("your current balance is " + balance.amount);
});
});
});
});
}
catch(err) {
res.end("Withdrawal error: " + err.message);
}
Note: The above code, although it looks like it will catch the exceptions, it will not.
Catching exceptions with callback hell adds a lot of pain, and i'm not sure if you will have the 'res' parameter
to respond to the user. If somebody like to fix this example... be my guest.
using wait.for:
var db = require("some-db-abstraction"), wait=require('wait.for');
function handleWithdrawal(req,res){
try {
var amount=req.param("amount");
sessiondata = wait.forMethod(db,"select","* from session where session_id=?",req.param("session_id"));
accountdata= wait.forMethod(db,"select","* from accounts where user_id=?",sessiondata.user_ID);
if (accountdata.balance < amount) throw new Error('insufficient funds');
wait.forMethod(db,"execute","withdrawal(?,?)",accountdata.ID,req.param("amount"));
res.write("withdrawal OK, amount: "+ req.param("amount"));
balance=wait.forMethod(db,"select","balance from accounts where account_id=?", accountdata.ID);
res.end("your current balance is " + balance.amount);
}
catch(err) {
res.end("Withdrawal error: " + err.message);
}
Note: Exceptions will be catched as expected.
db methods (db.select, db.execute) will be called with this=db
Your Code
In order to use wait.for, you'll have to STANDARDIZE YOUR CALLBACKS to function(err,data)
If you STANDARDIZE YOUR CALLBACKS, your code might look like:
var wait = require('wait.for');
//run in a Fiber
function process() {
var a = [1, 2, 3, 4, 5];
var result = 0;
a.forEach(function(i) {
// I/O operation
var record = wait.for(findFromDb,i); //call & wait for async function findFromDb(i,callback)
if (!record) throw new Error('No record exist for ' + i);
result += record.value;
});
return result/a.length;
}
function inAFiber(){
console.log('result is: ',process());
}
// run the loop in a Fiber (keep node spinning)
wait.launchFiber(inAFiber);
see? closer to python and no callback hell
Here is the situation:
I am new to node.js, I have a 40MB file containing multilevel json file like:
[{},{},{}] This is an array of objects (~7000 objects). Each object has properties and a one of those properties is also an array of objects
I wrote a function to read the content of the file and iterate it. I succeeded to get what I wanted in terms of content but not usability. I thought that I wrote an async function that would allow node to serve other web requests while iterating the array but that is not the case. I would be very thankful if anyone can point me to what I've done wrong and how to rewrite it so I can have a non-blocking iteration. Here's the function that handles the situation:
function getContents(callback) {
fs.readFile(file, 'utf8', function (err, data) {
if (err) {
console.log('Error: ' + err);
return;
}
js = JSON.parse(data);
callback();
return;
});
}
getContents(iterateGlobalArr);
var count = 0;
function iterateGlobalArr() {
if (count < js.length) {
innerArr = js.nestedProp;
//iterate nutrients
innerArr.forEach(function(e, index) {
//some simple if condition here
});
var schema = {
//.....get props from forEach iteration
}
Model.create(schema, function(err, post) {
if(err) {
console.log('\ncreation error\n', err);
return;
}
if (!post) {
console.log('\nfailed to create post for schema:\n' + schema);
return;
}
});
count++;
process.nextTick(iterateGlobalArr);
}
else {
console.log("\nIteration finished");
next();
}
Just so it is clear how I've tested the above situation. I open two tabs one loading this iteration which takes some time and second with another node route which does not load until the iteration is over. So essentially I've written a blocking code but not sure how to re-factor it! I suspect that just because everything is happening in the callback I am unable to release the event loop to handle another request...
Your code is almost correct. What you are doing is inadvertently adding ALL the items to the very next tick... which still blocks.
The important piece of code is here:
Model.create(schema, function(err, post) {
if(err) {
console.log('\ncreation error\n', err);
return;
}
if (!post) {
console.log('\nfailed to create post for schema:\n' + schema);
return;
}
});
// add EVERYTHING to the very same next tick!
count++;
process.nextTick(iterateGlobalArr);
Let's say you are in tick A of the event loop when getContents() runs and count is 0. You enter iterateGlobalArr and you call Model.create. Because Model.create is async, it is returning immediately, causing process.nextTick() to add processing of item 1 to the next tick, let's say B. Then it calls iterateGlobalArr, which does the same thing, adding item 2 to the next tick, which is still B. Then item 3, and so on.
What you need to do is move the count increment and process.nextTick() into the callback of Model.create(). This will make sure the current item is processed before nextTick is invoked... which means next item is actually added to the next tick AFTER the model item has been created... which will give your app time to handle other things in between. The fixed version of iterateGlobalArr is here:
function iterateGlobalArr() {
if (count < js.length) {
innerArr = js.nestedProp;
//iterate nutrients
innerArr.forEach(function(e, index) {
//some simple if condition here
});
var schema = {
//.....get props from forEach iteration
}
Model.create(schema, function(err, post) {
// schedule our next item to be processed immediately.
count++;
process.nextTick(iterateGlobalArr);
// then move on to handling this result.
if(err) {
console.log('\ncreation error\n', err);
return;
}
if (!post) {
console.log('\nfailed to create post for schema:\n' + schema);
return;
}
});
}
else {
console.log("\nIteration finished");
next();
}
}
Note also that I would strongly suggest that you pass in your js and counter with each call to iterageGlobalArr, as it will make your iterateGlobalArr alot easier to debug, among other things, but that's another story.
Cheers!
Node is single-threaded so async will only help you if you are relying on another system/subsystem to do the work (a shell script, external database, web service etc). If you have to do the work in Node you are going to block while you do it.
It is possible to create one node process per core. This solution would result in only blocking one of the node processes and leave the rest to service your requests, but this feature is still listed as experimental http://nodejs.org/api/cluster.html.
A single instance of Node runs in a single thread. To take advantage
of multi-core systems the user will sometimes want to launch a cluster
of Node processes to handle the load.
The cluster module allows you to easily create child processes that
all share server ports.
I'm using Mongoose with Node.js and have the following code that will call the callback after all the save() calls has finished. However, I feel that this is a very dirty way of doing it and would like to see the proper way to get this done.
function setup(callback) {
// Clear the DB and load fixtures
Account.remove({}, addFixtureData);
function addFixtureData() {
// Load the fixtures
fs.readFile('./fixtures/account.json', 'utf8', function(err, data) {
if (err) { throw err; }
var jsonData = JSON.parse(data);
var count = 0;
jsonData.forEach(function(json) {
count++;
var account = new Account(json);
account.save(function(err) {
if (err) { throw err; }
if (--count == 0 && callback) callback();
});
});
});
}
}
You can clean up the code a bit by using a library like async or Step.
Also, I've written a small module that handles loading fixtures for you, so you just do:
var fixtures = require('./mongoose-fixtures');
fixtures.load('./fixtures/account.json', function(err) {
//Fixtures loaded, you're ready to go
};
Github:
https://github.com/powmedia/mongoose-fixtures
It will also load a directory of fixture files, or objects.
I did a talk about common asyncronous patterns (serial and parallel) and ways to solve them:
https://github.com/masylum/i-love-async
I hope its useful.
I've recently created simpler abstraction called wait.for to call async functions in sync mode (based on Fibers). It's at an early stage but works. It is at:
https://github.com/luciotato/waitfor
Using wait.for, you can call any standard nodejs async function, as if it were a sync function, without blocking node's event loop. You can code sequentially when you need it.
using wait.for your code will be:
//in a fiber
function setup(callback) {
// Clear the DB and load fixtures
wait.for(Account.remove,{});
// Load the fixtures
var data = wait.for(fs.readFile,'./fixtures/account.json', 'utf8');
var jsonData = JSON.parse(data);
jsonData.forEach(function(json) {
var account = new Account(json);
wait.forMethod(account,'save');
}
callback();
}
That's actually the proper way of doing it, more or less. What you're doing there is a parallel loop. You can abstract it into it's own "async parallel foreach" function if you want (and many do), but that's really the only way of doing a parallel loop.
Depending on what you intended, one thing that could be done differently is the error handling. Because you're throwing, if there's a single error, that callback will never get executed (count won't be decremented). So it might be better to do:
account.save(function(err) {
if (err) return callback(err);
if (!--count) callback();
});
And handle the error in the callback. It's better node-convention-wise.
I would also change another thing to save you the trouble of incrementing count on every iteration:
var jsonData = JSON.parse(data)
, count = jsonData.length;
jsonData.forEach(function(json) {
var account = new Account(json);
account.save(function(err) {
if (err) return callback(err);
if (!--count) callback();
});
});
If you are already using underscore.js anywhere in your project, you can leverage the after method. You need to know how many async calls will be out there in advance, but aside from that it's a pretty elegant solution.