In addition, I think it's important that the postage stamps not be naively packaged as blobs containing FITS file contents; the overhead of that seems very high. At a minimum, lossless gzip compression might be used to substantially reduce the header overhead.

Agreed--I'm planning to simply gzip compress them for the purposes of this ticket since the proportion of header/image data is pretty large.

Kian-Tat Lim, can you take a look at this?

I generated an avro-serialized alert packet with 106 DiaSources, 9 DiaForcedSources, and gzip-compressed image cutouts.

Using ls -l for sizing, I obtain the following packet sizes:

Avro with schema included: 126KB
Avro with no schema: 82KB

I note that gzip-compressing the above packets yields 68KB/65KB, which may suggest an avenue of further exploration.

One bug that may have caused packets to be slightly small (or might not be exercised; I didn't check), but otherwise looks good. Sorry it took so long for me to look it over.

No change in the packet size after the review updates. I also added some more flexibility to the code to enable us to isolate how much of the packet comes from various components (e.g., the DIASource history, etc.).