将实时数据记录到 NoSQL DB 的最佳实践

Question

I got a Java web application which receives some real-time events and pushes them to the user interface layer. I want to log all the perceived events and since the volume of information will be huge, I prefer using a NoSQL db.

I have setup a mongodb for this purpose which inserts a document per event. The problem is that this approach (a disk access per event) slows down the whole process dramatically.

So, what approaches can I take in this situation? what options are available in mongodb for this (eg bulk inserting, async inserting, caching, ...)? would switching to some other NoSQL db implementation make a difference? what are the best practices here?

Answer 1

I have waited for some time to see other answers, but lose my patience. I have used MongoDB as a log storage for 3 projects (two for Java and one for C#). Basing on this I can figure out following important rules to organize logging:

1. Don't use indexes. If you mostly write then indexes cause performance degradation. If you need post-process log analyzes copy information to another database or collection. Unfortunately you cannot get rid of primary key _id - just leave it as is (GUID) or replace with auto-increment NumberLong .

2. Lower write-concern. MongoDB has rich options to control awareness of write operations. You can set matching between LogLevel and writing rules. For example DEBUG , INFO , WARN can go with WriteConcern.UNACKNOWLEDGED and ERROR , FATAL can be stored with WriteConcern.ACKNOWLEDGED . Such way you improve application performance by avoiding pause during low-priority messages writing. The same time you are sure that important messages (that are seldom) placed to storage.

3. Cache you collection instance. I mean avoid resolving Mongo's objects over getDB or getCollection each time when message arrives.

4. Minify amount of data passed by network. Restrict your message by minimal set of fields. Truncate too long stack trace. Look how Spring 3.x shortens full name of class swsmmaRequestMappingHandlerMapping instead of some.whatever.sub.main.minimal.agent.RequestMappingHandlerMapping