One case from a science collaboration member: they would like to (as I understand it) forward model supernova light curves on the direct images, not the difference images. To do so they need astrometric/photometric transformations between various pairs of images, which their past experience suggests can be more precise if the transformations are computed directly between the images rather than each being calibrated to an absolute frame.

Totally switching gears: there are almost certainly dwarf galaxies that will be discoverable in a single epoch of LSST imaging. The standard way to find them is by making appropriate color cuts on the source catalogs and find overdensities of main sequence turn-off stars or giants. Because there are strong incentives to be the first person to discover the dwarfs, it is very likely that users will not wait for a data release and will instead try to download and photometer images on their own if source catalogs are not available (this happened with DES).

Are we (or could we) store direct image Source catalogs on disk? This would enable both of those use cases without requiring additions to the PPDB. The number of Sources is much larger than the number of DIASources because most objects are not variable at LSST precision, so I am concerned about the implications for the PPDB sizing.

We could definitely just store catalogs in files on disk instead of in the database, and in many ways that might be the better solution. The one problem is that we have never specified a DPDD-level data product that is only delivered in this way; everything is generally a table in a database or an image (or attached to an image). Going that route would make the storage side easier but requires more work on the story for how users will access the data. I'll do some asking-around at all hands and see what that might look like on the technical side.

Sources, even for only a year (or 18 months) until the next DR comes out, are still over a PB including various overheads.  The addition of extra data products of this size needs careful consideration.

If we could somehow keep only the first Source at any given position, that would satisfy the dwarf galaxy use case without needing as much storage.

Pretty worried about erratic coverage quality if we literally did that, with some parts of the sky poorly covered because their first observation was on a bad night. But maybe that's OK given that uniform coverage is available relatively soon thereafter - the early discoveries might be misleading because of non-uniform sky coverage, but still interesting individually, and then the picture would be clarified at the next DR.

Regarding Colin Slater's initial use case of forward-modeling lightcurves, I would be more than happy to let that be in user-generated processing--unlike the dwarf galaxy case there are a finite number of supernovae (and hence images) you'd need to do it for. It's not even clear to me that there's urgency to do it before the Data Release, which would yield the highest-precision results anyway.

That's a fair point on the data volume; I had not looked up the numbers myself. I could imagine another minimal solution would be to store ~14 or 30 days of Sources. (One could reasonably argue that if you really need Sources from 6 months ago, you should just wait for the DR.) There might also be diagnostic merit in having that data, at least in the early stages of the survey. 

Repeating a point from the SST meeting today: this question is also tightly bound with the "what payload are we actually running on new data in LOY1?" question. For sky regions where we don't have templates yet, will there be any catalog data products at all, then, or just calibrated images (that we only keep around for 30 days before dropping them and leaving only lossy-compressed versions)? It's clear that there would be even more demand for Source-like data in some of these cases?

Some of this was already discussed in DMTN-107, Options for Alert Production in LSST Operations Year 1.

This was raised at the 2021 PCW today, so I wanted to chime in that for some science cases (e.g., searching for Solar System objects where there are no templates) even severely paired down source catalogs (just ra, dec, psfMag) would be extremely useful (and even if delivered as a series of FITS files). More generally, if the volume of the data is an issue, I suspect a subset of columns may satisfy a number of user cases discussed here.

I'll add this to a DM-SST meeting in September to discuss

Just in case I miss DM-SST meeting - it seems to me that it would be very useful for variety of reasons to have Source catalogs for the last k months (k~1) available in some form. As for the data volume, we can pair down the columns as Mario already said, and also play rounding tricks, similar to those that were used with early versions of the SDSS quasar catalog (e.g., don't store mag=12.3456789 but rather 12.35 or 12.346).