line_old

line

Data

Data is a first class citizen, Line's simple scheduling borrows from Hadoop's philosophy of moving compute to the data instead of the other way around.

• Dataset: a git repository
• Version: a git commit
• Replica: a clone of a Dataset on different specific node (no 2 replicas of the dataset are on the same node)
• Checkout: a local (mutable) working copy of specific Dataset Version

replicas eventually become in the same state, tasks require a replica at specific version

• Worker: a process that is responsible for:
  • reporting replica status
  • reporting alloc status

Tasks

The main purpose of Line is to run a container that take a specific Dataset Checkout as input and produce one or more versions of the dataset as output.

• Task: a planned container that takes N inputs and M outputs
• Input/Output: a local checkout of a certain dataset version, inputs are read-only. Outputs are writeable and committed when the task is finished.

A task can only be run if a dataset replica is present on the worker and checkouts for input and outputs can be placed on the node.

Hadoop like scheduling (compute to data)

• HDFS Arch guide: https://hadoop.apache.org/docs/r1.2.1/hdfs_design.html

Problems with state machine approach:

Current Challenges:

• how to deallocate on user/admin aborts of the state machines
• how to retry and deallocate on system failures: "States.Runtime" and "Lambda.Unkown"
• unaligned activity tokens
• very inefficient in a stable situation (infinite retries)