Repository navigation
ODB backward compatibility policy #11567
Replies: 6 comments 10 replies
|
I cannot recall that I've ever had to do more than open in a slightly different version. My most common use-case of ODB files is for caching between stages (i.e. not re-running synthesis) which would be fine with a short window in the tool. |
|
I would argue that support needs to be there to be able to open an ODB file for atleast 2-3 years, since the design -> test cycle can be very long. Recompiling OpenROAD with old versions is not viable (new OSs, compilers, etc might make it impossible to go back and finding the versions would be just as painful). |
|
I've heard 3mo (@oharboe ) 2-3yr (@gadfort), and 5yr (@QuantamHD). I've also heard just live with it as the "it always opens" user experience is more important than the developer annoyance. In the later case we could refactor a bit to make it less ugly but not ultimately less complex. |
|
One reframe: I think the trade-off is smaller than it looks, because dropping old formats from master doesn't have to be irreversible. #11484 doesn't need to be merged to be useful. It shows the legacy read paths can live outside mainline and be brought back if needed. So my suggestion: let maintainers drop formats once they feel the maintenance cost outweighs the benefit. If someone later hits a real need to open a very old .odb, #11484 (or a maintained version of it, per @gadfort's point about not keeping a pile of patches) is the fallback. Either way, expect some debugging with very old files. There's no regression suite that checks an old .odb behaves the same after loading. |
|
Thanks everyone for the input. Most votes went to B: native support with a finite window, and the discussion is settling on 3 years as enough for existing workflows. The key requirement is that out-of-support .odb files must always have a way to be upgraded, and the window shall be designed around that. |
Uh oh!
There was an error while loading. Please reload this page.
OpenROAD currently reads every ODB schema back to 57 (about 4 to 5 years). Each schema change adds another
if version < Xbranch to deserialization, so the code only grows. #11484 proposes a standaloneodb-converterto move that code out of mainline. Since this affects anyone with older.odbfiles, we'd like community input before deciding.The target case is users who update OpenROAD but never regenerate their
.odbfiles.Options
A. Status quo: native support forever. Easiest for users, but the code keeps growing.
B. Native support with a long finite window (e.g. 5 years): nearly as easy for users, and the code growth is bounded.
C. Short window plus multi-hop upgrade: each version reads a sliding window (e.g. 57 to 100, then 80 to 120, then 100 to 150). Users step old files forward through older commits. This keeps mainline small, but it's a hassle for users.
D. Short window plus
odb-converter(#11484): the upgrade is one step, but the converter inherits all the old branches unless it expires too. The converter moves the compatibility code rather than removing it.Questions
.odbfiles older than a year? How old are they?Idea: separate legacy reader
This could apply to any of the options: keep
operator>>current-format only and move old layouts into a separatedbLegacyRead.cpp, dispatched by one check at the top. The main reader stays clean, at the cost of duplicating the field list for the big classes (_dbBlock,_dbDatabase,_dbChip,_dbTechLayer). Thoughts?6 votes ·
All reactions