Comment 5 for bug 474869

This would hint that we might be having difficulties with our 'length estimation' code for a bzr-search index. We are putting data into a zlib compressor, and trying to estimate how long it will be. We estimate by using "zlib.flush(Z_SYNC)" in the stream, and I've seen in the past where the length with a few/bunch of Z_SYNC was actually *smaller* than the length of doing a zlib.compress() of the full data.

I'm guessing something about your data is provoking zlib. Or something about how bzr-search is grabbing terms, etc.

the 3976 is because we use 4096 byte pages, and reserve a 120 byte header for the first page (4096-120 = 3976).