Mongoose / MongoDB performance enhancements and tweaks (Part 3)

bush_doing_it_wrong_1

Holy crap. MongoDB native drivers are SO much faster than updates through mongoose’s ORM.  Initially, when I set out on this quest to enhance mongoDB performance with node.js, I thought that modifying my queries and limiting the number of results returned would be sufficient to scale. I was wrong (thanks Bush for the imagery). It turns out that the overhead that mongoose adds for wrapping mongoDB documents within mongoose’s ORM is tremendous. Well, I should say tremendous for our use case.

In my previous two posts of tweaking mongoDB / mongoose for performance enhancements (Part 1 and Part 2), I discussed optimization of queries or making simple writes instead of reads. These were worthwhile improvements and the speed difference eventually added up to significant chunks, but I had no idea moving to the native driver would give me these types of improvements (See below).

Example #1: ~400 streams, insertion times.
These numbers are from after I made the initial tweaks after Part 1. Unfortunately, I don’t really have that good of a printout from mongotop, but this kind of gives you an idea. Look at the write times for streams and packets, flowing at a rate of ~400 streams. This is for 400 sources of packets, which all gets written and persisted. Here you can see the write time to streams is @ 193ms / 400 streams or 48.25 ms / 100 streams. Likewise, packet writing is 7.25 ms / 100 streams. (You can mostly ignore read time, these are used for data aggregates and computing analytics). Compare these with the results below:

ns total read write
streams 193ms 0ms 193ms
packets 30ms 1ms 29ms
devices 9ms 9ms 0ms

Example 2: ~1000 streams, insertion times.
You can see here that write time has dropped significantly. Writes to the packets collection is hovering at around 1.7 ms / 1000 streams, and writes to the streams collection hovers at around 7.6 ms / 100 streams. Respectively, that’s a 425% and a 635% improvement in query write times to the packets collection and streams collection. And don’t forget, I had already begun the optimizations to mongoose. Even after the tweaks I made in Part 2, these numbers still represent a better than 100% improvement to query times. Huge, right?

ns total read write
packets 186ms 169ms 17ms
devices 161ms 159ms 2ms
streams 97ms 21ms 76ms

I knew using the mongoDB native drivers would be faster, but I hadn’t guessed that they would be this much faster.  

To make these changes, I updated mongoose to the latest version 3.8.14, which enables queries to be made using the native mongoDB driver released by 10gen (github here: https://github.com/mongodb/node-mongodb-native) via Model.Collection methods.  These in turn call methods defined in node_modules/mongodb/lib/mongodb/collection/core.js, which essentially just execute raw commands in mongo. Using these native commands, one can take advantage of things like bulk inserts.

I still like mongoose, because it helps instantiate the same object whenever you need to create and save something. If something isn’t defined in the mongoose.Schema, that object won’t get persisted to mongoDB either. Furthermore, it can still be tuned to be semi-quick, so it all depends on the use case. It just so happens that when you’re inserting raw json into mongoDB or don’t need the validation and other middleware that mongoose provides, you can use the mongoDB native drivers while still using mongoose for the good stuff. That’s cool.

Here’s what the new improvements look like:

    var Stream = mongoose.model('Stream');
    async.waterfall([
        //Native mongoDB update returns # docs updated, update by default updates 1 document:
        function query1 (callback) {
            Stream.collection.update(query1, query1set, {safe: true}, function(err, writeResult) {
                if (err) throw err;
                if (writeResult == 1) {
                    callback('Found and updated call @ Query1');
                } else {
                    callback(null);
                }
            });
        },
        function(callback) {
            Stream.collection.update(query2, query2set, {safe: true}, function(err, writeResult) {
                if (err) throw err;
                if (writeResult == 1) {
                    callback('Found and updated stream @ Query2');
                } else {
                    pushNewStream(packet, cb);
                    callback('No stream found.  Pushing new stream.');
                }
            });

        }
    ], function(err, results) {});
Advertisements

Mongoose / MongoDB speed improvements (Part 2)

In a previous post MongoDB performance enhancements and tweaks, I described some techniques I’ve found to speed up mongoose inserts and updates on large volume performance statistic data. I was still seeing performance bottlenecks in MongoDB, especially after running a node cluster for our data analytics. Since node is now spread to multiple cores (scaling “horizontally”, to be described in another post), the writes generally come to MongoDB much faster. The main question for me was whether it was mongoose, the node-mongoDB driver or mongo per se slowing me down.

The problem:

When we get performance data, we start tracking it by the creation of a “stream” object. The stream object is first created with a unique identifier, then subsequent packets that come in update the stream. The streams get the highest packet value of the incoming packet and update their timestamp with the packet’s timestamp. Later on when we stop seeing packets flow for a particular stream, we time it out so analytics can be computed for all packets that came in from the beginning of the stream to the end of the stream.

My initial implementation made a series of read queries using mongoose find(query), returned all potential matching streams, then updated the matching stream. The source code looked something like this.

function updateStream(packet) {
   var stream = mongoose.model('Stream');
   var query1 = {
        $and: [{
            'from.ID': packet.fromID
        }, {
            'to.ID': packet.toID
        }, {
            'stream_ended.from': false
        }, {
            'from.IP_ADDRESS': packet.IP
        }, {
            'from.highestPacketNum': {
                $lt: packet.highestPacketNum
            }
        }]
    };

   //Since streams are bilateral, we have to do two query reads in order to find the matching stream
   var query2 = {
        $and: [{
            'to.ID': packet.toID
        }, {
            'from.ID': packet.fromID
        }, {
            'stream_ended.to': false
        }, {
            'to.IP_ADDRESS': IP
        }, {
            'to.highestPacketNum': {
                $lt: packet.highestPacketNum
            }
        }]
    };  

   async.waterfall([
      function (callback) {
         stream.find(query1).exec(function(err, calls) { 
            if (calls.length > 1) {  //throw error, there should only be 1 stream
               throw new Error('There should only be one stream with this unique identifier');
            } else if (calls.length == 1) {
               //update calls[0]
               calls[0].save(cb);
               callback(null);
            } else {
               callback('Query1 yielded no results');  //continue down the waterfall
            }
         });
      },
      function (callback) {
         stream.find(query2).exec(function(err, calls) { 
            if (calls.length > 1) {  //throw error, there should only be 1 stream
               throw new Error('There should only be one stream with this unique identifier');
            } else if (calls.length == 1) {
               //update calls[0]
               calls[0].save(cb);
               callback(null);
            } else {
               callback('Query2 yielded no results');  //continue down the waterfall
            }
         });
      }
   ], cb);
}

You can see that this sourcecode was highly inefficient because mongoose was returning all possible matches of the stream. This was based on a limitation by our packet simulator, which at the time did not spoof unique IDs in the packets that it would send. At this point in time, we were capped at around 250 simultaneous streams, running in 1 node process.

Result: ~250 simultaneous streams of data

Improvement #1: Limit the search to the first object found, update it and persist it.

Essentially the source code remained the same, but the mongoose.find queries changed from find(query1).exec to find(query1).limit(1).exec(). With this change, we saw an improvement of around 50%, since mongoose would return after finding the first match. At this point, the blocker shifted back to node.js, and I noticed that at a certain point, the event queue would block up with handling too many events. AT the time, node.js was responsible for doing aggregation, decoding the packets, invoking mongoose and storing them, and running our REST API as well as serving up static data. I saw my poor little node.js process pinged out at 100% trying to churn through all the data. One good thing I noticed though, is that even though node.js was capped out in terms of resources, it still continued to churn through the data, eventually processing everything given enough time.

Result: ~350 simultaneous streams of data, query time improved by about 50%

Improvement #2: Clustering node.js

This deserves a post in itself since it required breaking out the various services that the single node process was handling into multiple forked processes, each doing its own thing. Suffice it to say, I eventually decided to fork the stream processors into N instances, where N is the number of cores on the machine / 2, with basic load balancing. This caused node to write to mongoDB much faster, and caused delays which eventually bogged mongoDB down. Thus the pendulum swung back to mongoDB.

Result: ~400 simultaneous streams of data, query time remains the same, mongoDB topped out.

Improvement #3: mongoDB updates in place

Finally, I realized that there was no reason to actually get the stream object returned in our source code, since I was just making updates to it. I had also noticed that it was actually the read time in mongotop that was spiking when the number of streams increased. This was because the find() functions in mongoose return a mongoose wrapped mongoDB document, so the update does not happen in place. For simple updates without much logic, there is no point to getting the object back, even when using the .lean() option to get json back. Here was the update:

function updateStream(packet) {
    //queries remain the same ...

    async.waterfall([
        function(callback) {
            Stream.findOneAndUpdate(query1, {$set: { 'endTime': packet.timestamp, 'metadata.lastUpdated': new Date(), 'from.highestPacketNum': highestPacketNum }}, {new: false}).exec(function(err, doc) {
                if (doc) {
                    cb(err, doc);
                    callback('Found and updated stream @ Query1');
                } else {
                    callback(null);
                }
            });
        },

        //No matching yet with IP, so try with just swapped SSRCs and update that one
        function(callback) {
            Stream.findOneAndUpdate(query2, {$set: { 'to.IP_ADDRESS': IP, 'endTime': packet.timestamp, 'metadata.lastUpdated': new Date(), 'to.highestPacketNum': highestPacketNum }}, {new: false}).exec(function(err, doc) {
                if (doc) {
                    cb(err, doc);
                    callback('Found and updated stream @ Query2');
                } else {
                    createNewStream(packet, cb);
                    callback('No streams found.  Pushing new stream.');
                }
            });
        }
    ], function(err, results) {

    });
}

It turns out that after this improvement, I saw in mongotop that read time dropped to 0ms per second, and write time slightly spiked. However, this was by far the biggest improvement in overall query time.

Result: ~600 simultaneous streams of data, query time dropped by 100 – 200% (when including read time dropping to 0). This, combined with the stream processors running on multiple cores seemed to be a scalable solution that would significantly spike our capacity, but I noticed that at around 600 simultaneous streams of data, suddenly our stream creation would spike, and continue increasing.

Improvement #4: MongoDB upgrade to 2.6.x and query improvement using $max field update operator

For all you readers who like mysteries, can you guess what was causing the stream creation to spike? I spent a long time thinking about it. For one, mongotop didn’t seem to be topped out in terms of the total query time on the streams collection. I noticed spikes of up to 400 or so ms for total query time, but it seemed by and large fine. Node.js was running comfortably at around 40% – 50% cpu usage per core on each of the four cores. So if everything seemed fine, why was stream creation spiking?

The answer, it turns out, was a race condition caused by the processing of the second packet of the simultaneous stream before the first packet could be used to instantiate a new stream object. At a certain point, when enough simultaneous streams were incoming, the delay in creation of the new stream would eclipse the duration between the first and second packets of that same stream. Hence, everything after this point created a new stream.

I thought for a while about a solution, but I got hung up either on choosing a synchronous processing of the incoming packets, which would significantly decrease our throughput, or use a “store and forward” approach where a reconciliation process would go back and match up streams. To be fair, I still have this problem, but I was able to reduce its occurrence to a significant extent. Because there’s no guarantee that we would be handling the packets in a synchronous order, I updated our query to make use of the $max field update operator, which would only update the stream highest packet number if a packet with the same IDs and higher packet number came in. This in turn let us reduce the query time because I no longer had to query to find a stream with a lower packet number than the incoming packet. After this update, I noticed that the reduced query time significantly reduced the total query time on the collection and at the same time helped the race condition issue.

function updateStream(packet) {
    var query1 = {
        $and: [{
            'from.ID': packet.fromID
        }, {
            'to.ID': packet.toID
        }, {
            'stream_ended.from': false
        }, {
            'from.IP_ADDRESS': packet.IP
        }]
    };

    var query2 = {
        $and: [{
            'to.ID': packet.fromID
        }, {
            'from.ID': packet.toID
        }, {
            'stream_ended.to': false
        }]
    };

    async.waterfall([

        //First see if there is a call object
        function(callback) {
            Call.findOneAndUpdate(query1, {
                $set: { 'endTime': packet.timestamp, 
                        'metadata.lastUpdated': new Date()
                    },
                $max: {
                    'from.highestPacketNum': highestPacketNum
                }
            }, {new: false}).exec(function(err, doc) {
                if (doc) {
                    cb(err, doc);
                    callback('Found and updated stream @ Query1');
                } else {
                    callback(null);
                }
            });
        },

        function(callback) {
            Call.findOneAndUpdate(query2, {
                $set: { 'to.IP_ADDRESS': packet.IP, 
                        'endTime': packet.timestamp, 
                        'metadata.lastUpdated': new Date()
                    },
                $max: {
                    'to.highestPacketNum': highestPacketNum 
                }
            }, {new: false}).exec(function(err, doc) {
                if (doc) {
                    cb(err, doc);
                    callback('Found and updated stream @ Query3');
                } else {
                    pushNewStream(packet, cb);
                    callback('No stream found.  Pushing new stream.');
                }
            });
        }
    ], function(err, results) {

    });
    return true;
}

Note that the threshold for the race condition is just higher with this approach and not completely solved. If enough streams come in, I’ll still eventually get the race condition where the stream is not instantiated before the second packet is being processed. I’m still not quite sure what the best solution is here, but as with everything, improvement is an incremental process.

Result: ~1000 simultaneous streams, query time dropped by 100%, 4 cores running at 40% – 50% cpu.

MongoDB performance enhancements and tweaks

MongoDB performance enhancements and tweaks

In my travails in building and my work on a real time analytics engine, I’ve formed some opinions on how well mongoDB is suited for scalability and how to tweak queries and my node.js code to extract some extra performance. Here are some of my findings, from several standpoints (mongoDB itself, optimizations to the mongoose driver for Node, and node.js itself).

Mongoose Driver
1. Query optimization

A. Instead of using Model.findOne or Model.find and iterating, try to use Model.find().limit() – I encountered a several factor speed up when doing this. This is talked about in several other places online.

B. If you have excess CPU, you can return a bigger chunk of documents and process them using your server instead and free up some cycles for MongoDB.

Improvement: Large (saw peaks of 1500ms for reads in one collection using mongotop. Afterwards, saw this drop to 200ms)

Example:

//Before:
Collection.findOne(query3, function(err, doc) {
  //Returns 1 mongoose document
});

//After
Collection.find(query3).limit(1).exec(function(err, docs) {
  //returns an array of mongoose documents            
});

See these links for some more information: Checking if a document exists – MongoDB slow findOne vs find

2. Use lean()
According to the docs, if you query a collection with lean(), plain javascript objects are returned and not mongoose.Document. I’ve found that in many instances, where I was just reading the data and presenting it to the user via REST or a visual interface, there was no need for the mongoose document because there was no manipulation after the read query.

Additionally, for relational data, if you have for instance a Schema that contains an array of refs (e.g. friends: [{ type: mongoose.ObjectId, ref: ‘User’}]), and you only need to return the first N number of friends to the user, you can use lean() to modify the returned javascript objects and then do population instead of populating the entire array of friends.

Improvement: Large (depending on how much data is returned)

Example:

//Before:
User.find(query, function(err, users) {
  //Users will be mongoose Documents. Hence you can't add fields outside the Schema (unless you have an { type: Any } object
  var options = {
     path: 'friends',
     model: 'User',
     select: 'first last'
  };
  Users.populate('friends', options, function(err, populated)) //will populate ALL friends in the array
});

//After
var query = new Query().lean();
User.find(query, function(err, users) {
  //Users will be javascript objects. Now you can go outside the schema and return data in line with what you need
  users.forEach(function(user) {
     user.friends = friends.splice(0, 10);  //take the first ten friends returned, or whatever
  });
  var options = {
     path: 'friends',
     model: 'User',
     select: 'first last'
  };
  Users.populate('friends', options, function(err, populated)) //now Model.populate populates a potentially much smaller array
});

Results (Example on my node.js server using mongoTop):
Load (ms)
Seconds No Lean() Lean()
5 561 524
10 371 303
15 310 295
20 573 563
25 292 291
30 302 291
35 544 520
40 316 307
45 289 286
50 537 503
Average 409.5 388.3
% improvement 0.051770452 = 5.177%

3. Keep mongoDB “warm”.
MongoDB implements pretty good caching. This can be evidenced by running a query several times in quick succession. When this occurs, my experience has been that the query time decreases (sometimes dramatically so). For instance, a query can go from 50ms to 10ms after running twice. We have one collection that is constantly queried – about 500 times per second for reads and also 500 times per second for writes. Keeping this collection “warm”, i.e. running the query that will be called at some point in the future, can help keep the call responsive when Mongo starts to slow down.

Improvement: Untested
Example:

function keepwarm() {
   setTimeout(function() {
      User.find(query);
      keepwarm();
   }, 500);
}

Mongo Native
1. Compound indexing
For heavy duty queries that run often, I decided to create compound indices using all the parameters that comprised the query. Even though intuitively, it didn’t jump out to me that indexing by timestamp for instance would make a difference, it does. According to the mongoDB documentation, if your query sorts based on timestamp (which ours did), indexing by timestamp can actually help.

Improvement: Large (depending on how large in documents your collection is and how efficiently mongoDB can make use of your indices)
Example:

//in mongo shell
db.collection.ensureIndex({'timestamp': 1, 'user': 1});

//in mongoose schema definition
Model.index({'timestamp': 1, 'user': 1});

Alternative? Aggregating documents into larger documents, such as time slices. Intuitively, that would mean that queries don’t have to traverse as large an index to reach the targeted documents. You may ask what the difference is between creating a compound index versus breaking the document down into aggregates like a day’s or hours slice. Here’s a few possibilities:

  1. A. MongoDB tries to match up queries with indices or compound indices, but there’s no guarantee that this match will occur. Supposedly, the algorithm used to determine which index to use is pretty good, but I question how good it is if for instance, the query you are using includes an additional parameter to search for. If MongoDB doesn’t see all parameters in the index, will it still know to use a compound index or a combination of compound indices?
  2. B. Using aggregates could actually be slower if it requires traversal of the document for the relevant flight data (which might not afford fast reads).
  3. C. If writes are very heavy for the aggregate (e.g. you use an aggregate document that is too large in scope), the constant reading and writing of the document may cause delays via mongoDB’s need to lock the collection/document.
  4. D. Aggregates could make indexing more difficult
  5. E. Aggregates could make aggregation/mapreduce more difficult because your document no longer represents a single instance of an “event” (or is not granular enough)

2. Use Mongotop to determine where your bottlenecks are.
Mongotop shows each collection in your database and the amount of time spent querying reads and writes. By default it updates every second. Bad things happen when the total query time jumps over a second. For instance, in Node, that means that the event queue will begin to block up because mongo is taking too long

Example:

 
//example output
                            ns       total        read       write		2014-07-31T17:02:06
              mean-dev.packets       282ms       282ms         0ms
             mean-dev.sessions         0ms         0ms         0ms
               mean-dev.series         0ms         0ms         0ms
              mean-dev.reduces         0ms         0ms         0ms
             mean-dev.projects         0ms         0ms         0ms

3. Use explain()… sparingly
I’ve found that explain is useful initially, because it will show you the number of scanned documents to reach the result of the query. However, when trying to optimize queries further, I found that it was not that useful. If I’ve already created my compound indices and MongoDB is using them, how can I extract further performance using explain() when explain() may already show a 0 – 1ms duration?

Example:

//in mongo shell
db.collection.find({
        $and: [{
            'from.ID': 956481854
        }, {
            'to.ID': 1038472857
        }, {
            'metadata.searchable': false
        }, {
            'to.IP_ADDRESS': '127.0.0.1'
        }, {
            'from.timestamp': {
                $lt: new Date(ISODate().getTime() - 1000 * 60 * 18)
            }
        }]
    }).explain()

4. For fast inserts for a collection of limited size, consider using a capped collection.

A capped collection in mongoDB is essentially a queue-like data structure that enforces first-in first-out. According to the mongoDB docs, capped collections maintain insertion order, so they’re perfect for time series. You just have to specify what the max size of the collection should be in bytes. I used an average based on: db.collection.stats(), where I found that each record was about 450 bytes in size.

To enforce this, you can run this in the mongoDB shell:

db.runCommand({"convertToCapped": "mycoll", size: 100000}); //size in bytes

See mongoDB docs here:

Node.js
1. Implement pacing for large updates.
I’ve found that in situations where there is a periodic update on a large subset of a collection while many updates are going on, the large update could cause the event queue in Node to backup as mongoDB tried to keep up. By throttling the number of updates that can go on based on total update time, I could adjust based on the load on the server currently. The philosophy is if node/mongoDB have extra cycles, we can dial up the pace of backfilling/updates a bit, whereas when node/mongoDB is overloaded, we can backoff.

Example:


//Runs periodically
    _aggregator.updateStatistics(undefined, updateStatisticsPace, function(result) {
          console.log('[AGGREGATOR] updateStatistics() complete.  Result: [Num Updated: %d, Duration: %d, Average (ms) per update: %d]', result.updated, result.duration, result.average);
          if (result.average < 5) {  //<5 ms, speed up by 10%
            updateStatisticsPace = Math.min(MAX_PACE, Math.floor(updateStatisticsPace * 1.1));    //MAX_PACE = all records updated
          } else if (result.average >= 5 && result.average < 10) { //5 < ms < 10, maintain pace
            updateStatisticsPace = Math.min(MAX_PACE, updateStatisticsPace);
          } else {  //>= 10ms, slow down by 2/3, to a min of 10
            updateStatisticsPace = Math.min(MAX_PACE, Math.max(updateStatisticsPace_min, Math.floor(updateStatisticsPace * .66)));
          }

          if (MAX_PACE === updateStatisticsPace) { console.log('[Aggregator] updateStatistics() - Max pace reached: ' + _count); }
          console.log('[AGGREGATOR] updateStatistics() Setting new pace: %d', updateStatisticsPace);
          callback(null, result)
    });

Dockumo – the crowd in the cloud

Screen Shot 2014-07-22 at 1.23.04 PM

In the past few months, I’ve been working on a project in my spare time called Dockumo. Dockumo is a web-based tool that lets users create “Articles” or basically blurbs of text, edit them and create new versions of them easily. When a user edits an article, the user is given the option to “Save” it or “Save as New Version.” Saving as a new version will generate a “Series” for the article, which will then start keeping track of all versions of the article. Pretty basic stuff. It’s my first foray into making consumer-facing software, including making better software tools for lawyers (my ultimate goal).

Dockumo main page (for text comparison)
Dockumo main page (for text comparison)

The utility here is that the user can see the article and versions through a timeline view and can easily compare any two versions (or any two articles for that matter) with a couple of clicks. I built this because comparison before required making two word documents, saving them, then comparing them and saving the result. Dockumo does this, but makes it much faster to do. Also, there’s no need to store lots of messy versions on your computer since it’s stored in the cloud. Furthermore, I built in some functionality to export the result to a text file, html doc or word doc (word was by far the most challenging). Exporting actually keeps track of the changes in “Track changes” format in word, which I implemented through a modification to the node.js officegen module.

An example of an article
An example of an article

Here’s what the timeline view looks like:

Timeline view of an article
Timeline view of an article

Finally, Dockumo lets users group articles together in “Journals.” Journals are a way to get related content together and then to share them with your friends or other users on the site. When you share a journal, it becomes a “collaboration” for all users who are invited. Those collaborators can add new versions of any articles grouped in the journal. The idea behind this is that users will be able to view and edit data together, while keeping track of new versions. There’s an option, for instance, to receive a notification whenever a collaborator edits a series. So for example, if five users are collaborators on a journal who are working on a report or a term paper, each one can make his or her modifications and save as a new version. As the document evolves through time, each user can see which other user made the change and how the article changed overtime through the comparison view.

A journal
A journal

What’s left?

When I started out making Docukmo, I only intended it to be a quick and easy way to diff two blurbs or text. I didn’t want feature creep to set in, but it inevitably did. Journals were not planned from the beginning, but I figured users would want a way to see relevant content. I ended up having to add lots of other elements as well, such as “friending” other users, user search, a notifications system, exporting to other formats, and emailing users who want updates, user permissions. All of these things added complexity and time. A good lesson for me though: When I thought I was able to launch the website, I was actually still about a month out. It wasn’t until I tried to launch did I realize that there were so many niggling things left undone. For instance, not allowing users to register under the same username or email address, or giving a password reset mechanism. Fortunately, I can carry over a lot of this code to any future projects down the line.

And feature-creep. I am probably the worst-offender of this concept ever. I always think to myself, “well if the software doesn’t do [X], then users won’t want to use it!” For me, this is an ever present temptation to continue adding feature upon feature to the source while never actually launching. Since launching, there have been other features that I really want to add (and probably will). For instance, I want to give users the option to make articles “public” and tag them with keywords so that the community (all users within the cloud) can search for a particular type of document (e.g. a cover letter or a thank you note), see an example of a it, and suggest modifications and rate existing articles. In other words, have a community of contributors who persistently improve cloud-based, crowd-sourced documents. That’s why I called the product “Dockumo”, or “doc” + “kumo” (“cloud” in Japanese).

Let me know what you guys think. I’m always happy to discuss improvements I can make or my technical/architectural decisions.

About page
About page

Real Time Analytics Engine

My current project at work is to architect a solution to capturing network packet data sent over UDP, aggregate it, analyze it and report it back up to our end users. At least, that’s the vision. It’s a big undertaking and there are several avenues of approach that we can take. To that end, I’ve started prototyping to create a test harness to see the write performance/read performance of various software at different levels in our software stack.

Our criteria are as follows (taken from a slide I created):

Our Criteria
Our Criteria

Based on these four requirements pillars, I’ve narrowed down the platform choices to the following:

SERVER

nodejsStrengths

      • Single-threaded, event loop model
      • Callbacks executed when request made (async), can handle 1000s more req than tomcat
      • No locks, no function in node directly performs I/O
      • Great at handling lots of small dynamic requests
      • Logo is cool

Weaknesses

      • Taking advantage of multi-core CPUs
      • Starting processes via child_process.fork()
      • APIs still in development
      • Not so good at handling large buffers

logo-white-big

Strengths

        • Allows writing code as single-threaded (Uses Distributed Event Bus, distributed peer-to-peer messaging system)
        • Can write “verticles” in any language
        • Built on JVM and scales over multiple cores w/o needing to fork multiple servers
        • Can be embedded in existing Java apps (more easily)

Weaknesses

        • Independent project (very new), maybe not well supported?
        • Verticles allow blocking of the event loop, which will cause blocking in other verticles. Distribution of work product makes it more likely to happen.
        • Multiple languages = debugging hell?
        • Is there overhead in scaling over multiple nodes?

PERSISTENCE

For our data, we are considering several NoSQL databases. Our data integrity is not that important, because our data is not make-or-break for a company. However, it is essential to be highly performant, with upwards of 10-100k, fixed-format data writes per second but much fewer reads.

mongo

Architecture

        • Maps in memory space directly to disk
        • B-tree indexing guarantee logarithmic performance

Data Storage

        • “Documents” akin to objects
        • Uses BSON (binary JSON)
        • Fields are key-value pairs
        • “Collections” akin to tables
        • No conversion necessary from object in data to object in programming language

Data-Replication

        • Mongo handles sharding for you
        • Uses a primary and secondary hierarchy (primary receives all writes)
        • Automatic failover

Strengths

        • Document BSON structure is very flexible, intuitive and appropriate for certain types of data
        • Easily query-able using mongo’s query language
        • Built-in MapReduce and aggregation
        • BSON maps easily onto JSON, makes it easy to consume in front end for charting/analytics/etc
        • Seems hip

Weaknesses

        • Questionable scalability and write performance for high volume bursts
        • Balanced B-Tree indices maybe not best choice for our type of data (more columnar/row based on timestamp)
        • NoSQL but our devs are familiar with SQL
        • High disk space/memory usage

 

cassandra
Architecture

        • Peer to peer distributed system, Cassandra partitions for you
        • No master node, so you can read-write anywhere

Data Storage

        • Row-oriented
        • Keyspace akin to database
        • Column family akin to table
        • Rows make up column families

Data Replication

        • Uses a “gossip” protocol
        • Data written to a commit log, then to an in-memory database (memtable) then to disk using a sorted-strings table
        • Customizable by user (allows tunable data redundancy)

Strengths

        • Highly scalable, adding new nodes give linear performance increases
        • No single point of failure
        • Read-write from anywhere
        • No need for caching software (handled by the database cluster)
        • Tunable data consistency (depends on use case, but can enforce transaction)
        • Flexible schema
        • Data compression built in (no perf penalty)

Weaknesses

        • Data modeling is difficult
        • Another query language to learn
        • Best stuff is used by Facebook, but perhaps not released to the public?

 

467px-Redis_Logo.svg

            TBD

 

        FRONT-END/REST LAYER

For our REST Layer/Front-end, I’ve built apps in JQuery, Angular, PHP, JSP and hands-down my favorite is Angular. So that seems like the obvious choice here.

AngularJS-large

Strengths

      • No DOM Manipulation! Can’t say how valuable this is …
      • Built in testing framework
      • Intuitive organization of MVC architecture (directives, services, controllers, html bindings)
      • Built by Google, so trustworthy software

Weaknesses

      • Higher level than JQuery, so DOM manipulation is unadvised and more difficult to do
      • Fewer 3rd party packages released for angular
      • Who knows if it will be the winning javascript framework out there
      • Learning curve

Finally, for the REST API, I’ve also pretty much decided on express (if we go with node.js):

express

Strengths

      • Lightweight
      • Easy integration into node.js
      • Flexible routing and callback structure, authorization & middleware

Weaknesses

      • Yet another javascript package
      • Can’t think of any, really (compared to what we are using in our other app – java spring

These are my thoughts so far. In the following posts, I’ll begin describing how I ended up prototyping our real time analytics engine, creating a test harness, and providing modularity so that we can make design decisions without too much downtime.