Chapter 5 - Worked Answers

1. Keep units throughout

0.4 CPU becomes400 millicores. Five replicas request2,000 millicores, or2 CPU. 64Mi is67,108,864 bytes per replica, totaling335,544,320 bytes, or320Mi. 64M is64,000,000 bytes per replica and320,000,000 bytes for five. The decimal/binary difference is15,544,320 bytes in the aggregate. These totals describe requested budget, not measured consumption or a proven host limit.

2. Identify the rejecting stage

The CPU case reaches ParseCPU and rejects an unparseable positive quantity. The memory case reaches ParseMemory, whose Mi branch requires a positive integer, so1.5Mi is invalid. The unknown replica field is rejected earlier by the YAML decoder's KnownFields mode. The actual fixture harness record includes the separate child commands and diagnostics, rather than treating a nonzero exit as sufficient evidence of any particular validation rule. No server is contacted by validate.

3. Compare actual paths

A node cannot provide the same host TCP8088 binding to two ordinary containers. The scheduler planner reserves node/protocol/port keys and rejects a second candidate. Immediate MemoryService task creation uses another helper which reserves but does not reject explicit conflicts. Therefore scheduler correctness alone does not protect this service path. The safe lab does not create a real shared-host collision; the source-confirmed gap is enough to reject the assumption. Dynamic public zero changes allocation behavior but cannot detect unrelated host processes outside orch's known tasks.

4. Trace caller data

Service resources are normalized and can be counted, but AgentTask does not transport them and ensureTask does not set ContainerSpec.Resources. The adapter's HostConfig code cannot enforce values it never receives. The real Docker contrast retained in the intake record shows zeros through the ordinary service path and the expected bytes/NanoCPUs through direct adapter invocation. To support isolation for Atlas itself, repeat an owned end-to-end service-path inspection after an independently reviewed implementation change. Passing a direct adapter test or modifying the book's baseline would not supply that evidence.

Return to Chapter 5.