How MongoDB Different Write Concern Values Affect Performance On A Single Node?


In the first post I talked about how indexes affect the write speed in MongoDB. In this second post I will share my findings on how different write concerns affect the write speed on a single node. Please refer to the first post for the setup related information.  A write concern controls the behavior of write operation and gives developers the choice to choose the value matching their requirements. For instance there are some documents which are not very important and if one of them get lost your business will not get screwed. For those you can choose less stricter value of write concern and for objects where you want don’t want your object to be lost you should choose stricter value of write concern. Let’s take a look at different write concern values available in Java driver. Please note in this experiment I used MongoDB java driver 2.7.2 instead of Spring MongoDB.

  1. Normal :  This is the default option where every write operation is fire and forget which means it just writes to the driver and return back. It does not wait for write to be available in server. So, if another thread tries to read the document just after the document has been written it  might found not find it. There is a very high probability of data loss with this option. I think this should not be considered in cases where data durability is important and you are only using single instance of MongoDB server. Even with replication you can loose data with this option (I will talk about in my future post).
  2. None : This is almost same as Normal with just one different that in Normal if network goes down or there is some other network issue you get an exception but with None you don’t get any exception if there are some network issues. This makes it highly unreliable.
  3. Safe : As suggested by name it is safer than the above two. The write operation waits for the MongoDB server to acknowledge the write but data is still not written to disk. With safe you will not face issue that when another thread tried to read the object you just wrote, the object was not found. So, it provides a guarantee that object once written will be found. That’s Good. But still you can loose data because data is not written to disk and if server died for some reason data will be lost.
  4. Journal Safe : Before we talk about this option. Lets first talk about what is Journaling in MongoDB. Journaling is a feature of MongDB where a write ahead log file of all the operations is maintained. In scenarios when MongoDB is not cleanly shutdown like using kill -9 command the data can be recovered from Journal files. By default data is written to journal files after every 100 milliseconds. You can change it to lie between 2 ms to 300 ms. With version 2.0 journaling is enable by default on 64 bit MongoDB servers. With Journal Safe write concern option your write will wait till the journal file is updated.
  5. Fysnc : With Fsync write concern the write operation waits till the data is not written to disk. This is the safest option on a Single node as only way you can loose data is when the hard disk crashes.

I have left the other values which are not applicable to single node but make more sense when replication is enable. I will cover them in future posts.

Test Case

The test case was very simple I will be doing 1 million writes with each of options except fsync and will find out the writes per second speed for each of the write concern values.

Document

The document is similar to the one used in first post. It is 2395 bytes.

{
"_id" : ObjectId("4eda74ef84ae8b2410f5fa8e"),
"age" : "27",
"lName" : "Gulati1",
"fName" : "Shekhar1"
"bio" : "I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. I am a Java Developer. ",
}

JUnit Test

The JUnit test case is shown below. In each test case it inserts one million records with a different value of write concern.

public class SingleNodeWriteConcernTests {

	private final static int ONE_MILLION = 1000000;

	private final Logger logger = Logger
			.getLogger(SingleNodeWriteConcernTests.class);

	@Test
	public void shouldInsertRecordsInNonCurrentMode() throws Exception {
		ServerAddress serverAddress = new ServerAddress("localhost", 27017);

		Mongo mongo = new Mongo(serverAddress);
		mongo.setWriteConcern(WriteConcern.NONE);
		runASingleTestCase(mongo, "NONE");

		mongo = new Mongo(serverAddress);
		mongo.setWriteConcern(WriteConcern.JOURNAL_SAFE);
		runASingleTestCase(mongo, "JOURNAL_SAFE");

		mongo = new Mongo(serverAddress);
		mongo.setWriteConcern(WriteConcern.NORMAL);
		runASingleTestCase(mongo, "NORMAL");

		mongo = new Mongo(serverAddress);
		mongo.setWriteConcern(WriteConcern.SAFE);
		runASingleTestCase(mongo, "SAFE");

	}

	private void runASingleTestCase(Mongo mongo, String name) throws Exception {
		DB db = mongo.getDB("play");
		DBCollection people = db.getCollection("people");
		if (db.collectionExists("people")) {
			people.drop();
		}
		insertRecords(mongo, name);

		mongo.dropDatabase("play");
	}

	private void insertRecords(Mongo mongo, final String name) throws Exception {

		DB db = mongo.getDB("play");
		final DBCollection collection = db.getCollection("people");
		collection.ensureIndex("fName");
		long startTime = System.currentTimeMillis();
		for (int i = 1; i <= ONE_MILLION; i++) {
			BasicDBObject obj = new BasicDBObject();
			Map<String, String> map = new HashMap<String, String>();
			map.put("fName", "Shekhar" + i);
			map.put("lName", "Gulati" + i);
			map.put("age", String.valueOf(i));
			map.put("bio", StringUtils.repeat("I am a Java Developer. ", 100));
			obj.putAll(map);
			collection.insert(obj);
		}
		long endTime = System.currentTimeMillis();
		double seconds = ((double) (endTime - startTime)) / (1000);
		double rate = ONE_MILLION / seconds;

		String message = String
				.format("WriteConcern %s inserted %d records in %.2f seconds at %.2f (rec/s)",
						name, ONE_MILLION, seconds, rate);
		logger.info(message);

	}

}

Results

As you might have also expected Normal and None are the fastest because of the way they work i.e. fire and forget. Safe writes takes 3.5 times more than Normal writes. With Journal safe value you come down to 24 documents per second which is very low. As you can see as you move towards more write safety you loose a lot on write speed. This is again a decision you have to make depending on your use case.

Can something be done to increase write speed in Safe and Journal Safe options?

The results shown above are based on records being inserted sequentially one at a time. I tried an experiment where in I divided 1 million records to a batch of 100,000 records each. And let 10 threads write 1 million record in parallel.  The write speed for Safe and Journal Safe increased but None and Normal decreased as shown below.

The write speed for Safe with 10 threads is 1.4 times the write speed with one thread and similarly write speed for Journal Safe is 10 times of the write speed with one thread. This is because while one thread is waiting other threads can work in parallel which allows to better utilize CPU.

About these ads

11 thoughts on “How MongoDB Different Write Concern Values Affect Performance On A Single Node?

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s