SArchive

Archival tool for scheduler job scripts and accompanying files.

Note that the master branch here may be running ahead of the latest release on crates.io. During development, we sometimes rely on dependencies that have not yet released a version with the features we use.

Minimum supported rustc

1.36

This version is what we test against in CI. We also test on - stable - nightly

If you do not have Rust, please see Rustup for installation instructions.

Usage

sarchive requires that the paths to the scheduler's main spool directory is specified. It also requires a cluster (the cluster name) to be set.

sarchive supports multiple schedulers, the one to use should also be specified on the command line. Right now, there is support for Slurm and Torque.

For Slurm, the directory to watch is the `

Furthermore, sarchive offers various backends. The basic backend writes a copy of the job scripts and associated files to a directory on a mounted filesystem. We also have limited support for sending job information to Elasticsearch or produce to a Kafka topic. We briefly discuss these backends below.

File archival

Activated using the file subcommand. Note that we do not support multiple subcommands at this moment.

For file archival, sarchive requires the path to the archive's top directory, i.e., where you want to store the backup scripts and accompanying files.

The archive can be further divided into subdirectories per - year: YYYY, by provinding --period=yearly - month: YYYYMM, by providing --period=monthly - day: YYYYMMDD, by providing --period=daily This allows for easily tarring old(er) directories you still wish to keep around, but probably no longer immediately need for user support. Each of these directories are also created upon file archival if they do not exist.

For example,

sarchive --cluster huppel -s /var/spool/slurm file --archive /var/backups/slurm/job-archive

Elasticsearch archival

If you want to maintain the job script archive on another machine and/or make it easily searchable, the Elasticsearch backend. The shipped data structure contains a timestamp along with the job script and potentially other relevant information (at the scheduler's discretion).

We do not yet support SSL/TLS or authentication with the ES backend.

For example,

sarchive --cluster huppel -s /var/spool/slurm elasticsearch --host myelastic.mydomain --index slurm-job-archive

Kafka archival

Similar to ES archival, no SSL/TLS support at this moment. Data is shipped in the same manner.

For example,

./sarchive --cluster huppel -l /var/log/sarchive.log -s /var/spool/slurm/ kafka --brokers mykafka.mydomain:9092 --topic slurm-job-archival

Features

• Multithreaded, watching one dir per thread, so no need for hierarchical watching.
• Separate processing thread to ensure swift draining of the inotify event queues.
• Clean log rotation when SIGHUP is received.
• Experimental support for clean termination on receipt of SIGTERM or SIGINT, where job events that have already been seen are processed, to minimise potential loss when restarting the service.
• Output to a file in a hierarchical directory structure
• Output to Elasticsearch
• Output to Kafka

RPMs

We provide a build script to generate an RPM using the cargo-rpm tool. You may tailor the spec file (listed under the .rpm directory) to fit your needs. The RPM includes a unit file so sarchive can be started as a service by systemd. This file should also be changed to fit your requirements and local configuration.