Currently, rosbag filter deserializes every message in a bag, in a single threaded process, with a fresh eval each time. This is all enormously expensive and slow:

Some possibilities for how to improve this situation:

• Lazy deserialization: Instead of deserializing every time, pass a proxy object which uses getattr to deserialize on demand and pass through accesses. This would prevent unnecessary deserialization expense for the use-cases which only filter on topic name or timestamp.
• Wrap the eval in a lambda, so that it needs to be compiled just once rather than on each use: https://stackoverflow.com/a/12467755/109517
• Use multiprocessing to parallelize the eval invocations—would have to check how much of a gain there is to be had here, particularly with the added cost of managing the work queue to maintain sequence.