How can you measure how good your MongoDB (or other databases with similar interface) performance is? Easy, you can benchmark it. A general way to solve this problem is to use a benchmark tool to generate queries with random contents under certain random distribution.

But sometimes you are not satisfied with the randomly generated queries, since you're not confident in how much these queries resemble your real workload.

The difficulty compounds when one MongoDB instance may host completely different types of databases that each have their own unique and complicated access patterns.

That is the reason we came up with Flashback, a MongoDB benchmark framework that allows us to benchmark with "real" queries. It is comprised of a set of scripts that fall into the 2 categories:

1. records the operations(ops) that occur during a stretch of time;
2. replays the recorded ops.

The two parts are not tied to each other and can be used independently for different purposes.

How do you know which ops are performed by MongoDB? There are a lot of ways to do this. But in Flashback, we record the ops by enabling MongoDB's profiling.

By setting the profile level to 2 (profile all ops), we'll be able to fetch the ops information detailed enough for future replay -- except for insert ops.

MongoDB does not log insertion details in the profile DB. However, if a MongoDB instance is working in a "replica set", we can capture insert information by reading the oplog.

Thus, we record the ops with the following steps:

1. The script starts multiple threads to pull the profiling results and oplog entries for collections and databases that we are interested in. Each thread works independently.
2. After fetching the entries, we'll merge the results from all sources to get a full picture of all operations.

With the ops being recorded, we also have a replayer to replay them in different ways:

• Replay ops with "best effort". The replayer diligently sends these ops to databases as fast as possible. This style can help us to measure the limits of databases. Please note to reduce the overhead for loading ops, we'll preload the ops to the memory and replay them as fast as possible. This potentially limits the number of ops played back per session to the available memory on the Replay host.
• Reply ops in accordance to their original timestamps, which allows us to imitate regular traffic.

The replay module is written in Go because Python doesn't do a good job in concurrent CPU intensive tasks.

• The "record" module is written in python. You'll need to have pymongo, mongodb's python driver installed.
• Set MongoDB profiling level to be 2, which captures all the ops.
• Run MongoDB in a replica set mode (even there is only one node), which allows us to access the oplog.

• If you are a first time user, please run cp config.py.example config.py.
• In config.py, modify it based on your need. Here are some notes:
  • We intentionally separate the servers for oplog retrieval and profiling results retrieval. As a good practice, it's better to pull oplog from secondaries. However profiling results must be pulled from the primary server.
  • duration_secs indicates the length for the recording.

After configuration, please simply run python record.py.

Go 1.4

$ go get github.com/ParsePlatform/flashback/cmd/flashback

Required options:

flashback \
    --style=[real|stress] \
    --ops_filename=<file_name> \ # Operations file, such as generated by the Record tool

For a full list of options:

flashback --help

pcap_converter is an experimental way to build a recorded ops file from a pcap of mongo traffic.

$ go get github.com/ParsePlatform/flashback/cmd/pcap_converter
$ tcpdump -i lo0 -w some_mongo_cap.pcap 'dst port 27017'
$ pcap_converter -f some_mongo_cap.pcap > ops_filename.json