LRU evict as a start flag

Ray Core
1

after ray start if I use ray.init(lru_evict=True,,...) I get an error stating that I cannot use this option when ray has already started
When connecting to an existing cluster, _internal_config must not be provided.

is there a way todo a ray start with the lru_evict=True option?

2

Hello, maybe you already have a ray runtime started, try ray stop and init the cluster again, here is the sample code which works on my laptop, hope this helps

Screen Shot 2021-05-11 at 10.24.48 PM
Screen Shot 2021-05-11 at 10.24.48 PM2582×980 279 KB

3

Checking the source I think I see that this flag is deprecated:

Set the internal config options for LRU eviction.

    if lru_evict:
        raise DeprecationWarning(
            "The lru_evict flag is deprecated as Ray natively "
            "supports object spilling. Please read "
            "https://docs.ray.io/en/master/memory-management.html#object-spilling "  # noqa
            "for more details.")
4

Yeah the flag is deprecated from 1.3 because the disk spilling is on by default. Theoretically, there should be no need for lru evict since now Ray manages distributed references & OOM (by disk spilling) properly.

5

Thank you for the details.
I am still using Ray 0.8, definitely upgrading to 1.3 will solve my problem.
However this flag does not exist in 0.8, it only exist as a Ray init() parameter. Can I use Ray start and have this option set (with the Ray start flags)?

6

When I looked at Ray 0.8’s code, it doesn’t look like lru_evict is supported from that version (it is probably supported from 0.8.1?)