JosepSampe
commented
4 months ago The issue that explains v1 and v2 is https://github.com/lithops-cloud/lithops/issues/1174
In short, the difference between v1 and v2 is the granularity of execution. In v1 the granularity is at the job level and in v2 the granularity is at the task level. Let me explain this with an example:
Lets assume that we have 8 cores in our machine, a function called my_func that takes 5 minutes to complete, and the next code:
def my_func():
time.sleep(300)
fexec = FunctionExecutor()
fexec.map(my_func, range(1))
fexec.map(my_func, range(7))
fexec.get_result()- v1: localhost v1 will initially run the first map until it completes and then proceed to run the second map. This means that during the execution of the first map, 7 cores will be free, and at the end the total execution time will be 10 minutes.
- v2: localhost v2 will run all 8 tasks at the same time, which means there will be no underutilized resources, and at the end of the script the total execution time will be 5 minutes.
So, v2 is more precise in what you will experience when running tasks in the cloud.
In addition, the recently added "Retry" feature will also not work as expected with version 1. Since if you have a 10 minute straggler, the rerty logic will invoke a new task, however this new task will not run until the straggler finish, which is not the expected behavior in this case.
tomwhite
I have only seen this on Linux CI (see https://github.com/cubed-dev/cubed/pull/471), and haven't been able to reproduce locally on a Mac.
The fix is to use v1 of the localhost backend - this should probably be made the default again until the problem is diagnosed and fixed.
@JosepSampe what is the difference between the two versions? I couldn't see it explained anywhere in GH issues but perhaps I missed it.
Also, I noticed that the documentation says the default is v1, and that v2 is beta: https://lithops-cloud.github.io/docs/source/compute_config/localhost.html#summary-of-configuration-keys-for-localhost.