Meadowrun

**Breaking changes**:

This release changes the format of the allocation tables. You must run `meadowrun-manage-ec2 uninstall && meadowrun-manage-ec2 install` after updating if you have used any previous versions of Meadowrun before.

Improvements:

- More run_map performance and reliability improvements
- Better support for running Meadowrun instances in a private subnet in EC2

Full Changelog: https://github.com/meadowdata/meadowrun/compare/v0.2.6...v0.2.8

More robustness improvements:
- Raise a special error explaining what's happening when we delete AMIs for old versions
- Address some cases where run_map would hang

Note: EC2 AMIs were not updated as only client code was changed.

**Full Changelog**: https://github.com/meadowdata/meadowrun/compare/v0.2.5...v0.2.6

New features:
- Specify a subnet_id or ami_id for EC2
- Add a max_num_task_attempts parameter to run_map
- Add the always_use_local flag to ContainerInterpreter and PreinstalledInterpreter

Robustness improvements:
- run_map will start using existing instances immediately rather than waiting until both existing instances and newly launched instances are available.
- Logging improvements: don't stream logs for run_map if there's more than one worker. Print out the host/path to the log where it would be useful. Print out remote exceptions in a nice format.
- Proactively terminate instances and deallocate jobs when the client process is interrupted. This means that we don't need the deallocate_jobs cron job to realize that a job was killed or never started.
- Include the "agent" (run_job_local) logging in the log files

Bug fixes:
- Make /var/meadowrun/machine_cache available on Kubernetes
- Poetry environment creation on Kubernetes was not working

**Full Changelog**: https://github.com/meadowdata/meadowrun/compare/v0.2.4...v0.2.5

Breaking API changes:
- Changed the machine cache folder from /meadowrun/machine_cache to /var/meadowrun/machine_cache so that it is consistent across container and non-container jobs. Added the `meadowrun.MACHINE_CACHE_FOLDER` variable so that users don't need to hardcode this folder.

Full Changelog: https://github.com/meadowdata/meadowrun/compare/v0.2.3...v0.2.4

Minor improvements and bug fixes:
- Improvements to reliability and scalability of `run_map`
- Fix a bug where conda environment specs that require cuda or other additional software was not working
- mirror_local can now take a single string for the `additional_python_paths` parameter

**Full Changelog**: https://github.com/meadowdata/meadowrun/compare/v0.2.2...v0.2.3

New features:
- Add support for pip, poetry, and conda environment syncing on Kubernetes.
- Add run_map_as_completed

**Full Changelog**: https://github.com/meadowdata/meadowrun/compare/v0.2.1...v0.2.2