2-platform local clone() exhausts memory

[khaeru]

Code sample or context

1. Clone a large (e.g. ENGAGE) message_ix.Scenario into a ‘local’ platform (i.e. using JDBCBackend with driver="hsqldb").
2. Clone from one local platform to another.

Expected result

Clone works.

Problem description

The process is killed for exceeding memory limits.

This appears to only occur under certain conditions:

• On my local machine. I have 16 GB of RAM and use jvmargs=["-Xmx12G"] for the JDBCBackend.
• When using message_data.model.transport. It does work when use a few commands on the terminal, like:

  from ixmp import Platform
  from message_ix import Scenario
  s, mp1 = Scenario.from_url("ixmp://local-1/…")
  mp2 = Platform("local-2")
  s.clone(platform=mp2)
  mp2.close_db()

Versions

Latest main.

Discussion

• The on-disk size of the "local-1" platform, containing 3 such large scenarios, is about 342 MiB, so each scenario is about 100 MiB compressed.
• 100 MiB of data should probably not take up >12 GiB (or >6 GiB) in memory.
• This may be because the Java code in ixmp_source loads all contents of the Scenario before performing the clone.
• If so, one response could be to clone item-wise.
• The responses could be:
  1. Modify the Java code to disable this.
  2. Reimplement cloning in Python.
• (2) will eventually be required in order to support backends other than JDBCBackend.
• Doing (2) would reduce the scope and ease maintenance/replacement of JDBCBackend.

[OFR-IIASA]

I can confirm the issue. I have tried to do the same and encountered that case (1) as described above doesnt. In addition I think is also worth mentioning that the issue of running out of memory is only encountered after a total of 70 minutes.