chore: Remove session scheduler related types no longer used

[jopemachine]

Overview

During the process of refactoring tasks of different characters, I discovered that these types are no longer in use.
It seems advisable to remove these types for maintenance.

Details

• Remove the following types no longer used.

  • AgentContext
    • Replaced with AgentRow in refactor: Split the session & kernel tables with ORM migration #576.
  • ExistingSession
    • Replaced with SessionRow in refactor: Split the session & kernel tables with ORM migration #576.
  • PendingSession
    • Replaced with SessionRow in refactor: Split the session & kernel tables with ORM migration #576.
  • SchedulingPredicate
    • Created in Job queueing backend.ai-manager#192, but I couldn't find where SchedulingPredicate is used.
  • StartTaskArgs
    • Created in Stabilization of multi-node multi-container sessions (Phase 3) backend.ai-manager#376. In that PR, there was code that used the StartTaskArgs type, but it is no longer present in version 20.09.

• Replace the KernelInfo type argument of assign_agent_for_kernel with KernelRow.
  The KernelInfo type is not actually used, and at the call site of assign_agent_for_kernel, a value of type KernelRow is passed to that parameter. You can see the kernel passed to the assign_agent_for_kernel is KernelRow type in the below code.

  backend.ai/src/ai/backend/manager/scheduler/dispatcher.py

  Lines 1027 to 1094 in 62eca72

  	kernel: KernelRow
  	for kernel in sess_ctx.kernels:
  	agent_alloc_ctx: AgentAllocationContext | None = None
  	try:
  	agent_id: Optional[AgentId] = None
  	agent: Optional[AgentRow] = kernel.agent_row
  	if agent is not None:
  	# Check the resource availability of the manually designated agent
  	result = (
  	await agent_db_sess.execute(
  	sa.select([
  	AgentRow.available_slots,
  	AgentRow.occupied_slots,
  	]).where(AgentRow.id == agent.id)
  	)
  	).fetchall()[0]
  	if result is None:
  	raise GenericBadRequest(f"No such agent exist in DB: {agent_id}")
  	available_slots, occupied_slots = result
  	for key in available_slots.keys():
  	if (
  	available_slots[key] - occupied_slots[key]
  	>= kernel.requested_slots[key]
  	):
  	continue
  	else:
  	raise InstanceNotAvailable(
  	extra_msg=(
  	f"The designated agent ({agent.id}) does not have "
  	f"the enough remaining capacity ({key}, "
  	f"requested: {sess_ctx.requested_slots[key]}, "
  	f"remaining: {available_slots[key] - occupied_slots[key]})."
  	),
  	)
  	agent_id = agent.id
  	else:
  	# Each kernel may have different images and different architectures
  	compatible_candidate_agents = [
  	ag for ag in candidate_agents if ag.architecture == kernel.architecture
  	]
  	if not candidate_agents:
  	raise InstanceNotAvailable(
  	extra_msg="No agents are available for scheduling"
  	)
  	if not compatible_candidate_agents:
  	raise InstanceNotAvailable(
  	extra_msg=(
  	"No agents found to be compatible with the image architecture "
  	f"(image: {kernel.image_ref}, "
  	f"arch: {kernel.architecture})"
  	),
  	)
  	available_candidate_agents = await self._filter_agent_by_container_limit(
  	compatible_candidate_agents
  	)
  	if not available_candidate_agents:
  	raise InstanceNotAvailable(
  	extra_msg=(
  	"No agents found to be available because all agents have"
  	" reached the hard limit of the number of containers."
  	),
  	)
  	# Let the scheduler check the resource availability and decide the target agent
  	agent_id = scheduler.assign_agent_for_kernel(
  	available_candidate_agents,
  	kernel,

Note

Unlike other types, I couldn't find any evidence that the SchedulingPredicate type was used in the past, so there might be some misunderstanding on my part.
If that's the case, please let me know. If it is a useful type somewhere, we can fix it by replacing the removed PendingSession type with SessionRow.

---

Checklist: (if applicable)

• Milestone metadata specifying the target backport version

[jopemachine]

• chore: Remove session scheduler related types no longer used #2705 👈
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

Join @jopemachine and the rest of your teammates on Graphite