Asklv
Asklv
> Do I need to increase the test coverage for `JobSet` validation? I think you are right, we can refer to @tenzen-y PR , https://github.com/kubeflow/trainer/pull/2555.
Hi Contributor @Okabe-Rintarou-0 , I also find this problem and would like to confirm if the PR/issue is still active? cc @andreyvelich
Check tests: ```rs test storage::storage_manager::test::concurrency::test_concurrency_aligned ... ok test storage::storage_manager::test::concurrency::test_concurrency_unaligned ... ok test new_storage::storage_manager::tests::storage_real_workload_test ... ok test result: ok. 189 passed; 0 failed; 3 ignored; 0 measured; 0 filtered out; finished...
Version changed: 20M ```bash daten-server:~/data$ dd if=/dev/zero bs=1M count=20 | tr '\0' '1' > datenlord_cache/new_storage_20m_test_version10.txt 20+0 records in 20+0 records out 20971520 bytes (21 MB, 20 MiB) copied, 10.5831 s,...
> FileHandle.access() can be remove OK, file level version PR will solve the block mismatch issue, if remove the access function now, the fastest will failed. > Current fstest is...
truncate 20M: version change: 10 write: version change: 556
Because we want to use singleton task manager, this manager will create a task in current runtime context, it is isolated in tokio::test. Reference issues: https://github.com/tokio-rs/tokio/issues/2374
The current debugging result should be pointing to the runtime problem, that Lazy task manager should use the default context runtime in the new tokio::spawn, and then in the test...
The Issue already mentions the need to introduce documentation similar to the `TrainerPipelineFramework` in [issue](https://github.com/kubeflow/trainer/issues/2458#issuecomment-2698386151), which I've done a simple fine-tuning of, but I'm not sure if there are any...
> Hi @IRONICBo, did you get a chance to address the remaining comments ? Having this documentation would be super helpful to understand the KF Trainer architecture. @andreyvelich Sorry for...