What is MongoDB's fsync for?

Question

I've been using MongoDB for some time, and saw that fsync waits for data to be flushed to disk. Ok, so i thought it was the solution for safety of the data.

It worked well by takes long, longer than SQL alternative. Then I saw that I can put the syncdelay to 0 , then speed came back, but I thought how it would be in the future with many many concurrent requests. So I removed fsync option from the updates and inserts and removed the syncdelay configuration option.

To test if the data was being written I quickly checked Rockmongo after I made an update and the data was actually there, super fast!

So really, what is fsync for if it makes the writes slow and without it the writes happen, and fast anyway?

Answer 1

Per Mongo documentation:

The primary use of fsync is to flush and lock the database for backups.

also

The fsync operation blocks all other write operations while its running.

The blocking appears to be the reason.

Answer 2

fsync is technically an admin command that forces a flush of all data to disk. You shouldn't have to use it in your code, not normally at least. It's used to lock the database for backups and so on.

Data safety in MongoDB comes from replication/sharding/journaling, not from forcing writes. That kind of defeats the purpose of the thing.

The Java driver wraps this 'write-and-sync' concept in the WriteConcern class, which I've never really liked much. You shouldn't have to decide which part of your data is more or less important, but rather just trust the tool to do its job.

Also, if you set syncdelay to zero make sure you turn journaling off. See this .

Answer 3

As the other answers have said, the fsync command forces a flush and is normally used right before you lock data files for a point in time snapshot.

There is an "fsync" write concern option on getLastError that will wait to return into all pending data has been flushed to disk. You normally wouldn't use this though, the "j" option (which returns as soon as journaling has happened) is much faster to return and still ensures durable writes. You can pass either through an update/insert command as the safe option in your driver of choice to let it automatically run the getLastError command for you.