High rates of corrupt bags with rosbag python api?

I have a python action server where the callback creates a `rosbag.Bag` and then callback subscribing to various messages then write to that bag, and when the action is over the bag is closed and the callbacks then return without attempting to write. It appears that about 30%-50% of the bags that come out of this say they need to be reindexed, and then after `rosbag reindex` a huge amount of the messages are lost- what ought to be hundreds of messages are now 10s or even zero- sometimes the reindexed bag is the same size as orig, other times smaller.