GPUs, TPUs, and a snapshot schedule nobody had tested
Choose accelerators by framework fit and workload shape, and a backup that's never been restored is an unverified assumption.
Speaker notes
- GPUs offer the broadest framework compatibility for general parallel compute, TPUs are Google's own custom silicon, purpose-built for large-scale ML training and inference, outperforming specifically on the matrix-multiplication-heavy patterns they're designed around.
- Accelerator availability is far more region and zone-specific than standard machine types, a ZONE_RESOURCE_POOL_EXHAUSTED error is often an availability problem, not a quota problem, check the regional availability table before writing the create command. A snapshot restores data to a disk; an image creates new bootable instances, they solve different problems even though you can convert between them.
- Tell the snapshot incident: a daily snapshot schedule 'succeeded' for a year before a real restore attempt revealed it had been backing up an empty placeholder disk for five months after an undetected machine-type migration. A recurring, verified test restore is the only thing that actually confirms a backup works.
Deck map
01
GCP foundations, fully current
02
The exam itself, and how this deck maps to it
03
The project is the real unit of isolation
04
Organization, folders, projects, and now: standalone orgs
05
Org policies now ship with a dry-run mode
06
Five more day-zero considerations
07
Three metadata mechanisms, one exam trap
08
Cloud Identity: manual and SCIM-automated
09
Cloud Asset Inventory, Gemini Cloud Assist, and Workforce Identity Federation
10
A budget alerts, it never caps
11
gcloud, Cloud Shell, and the ADC trap
12
Five infrastructure tools, five different jobs
13
Gemini CLI sunset, Antigravity, and Application Design Center
14
Principal, permission, role, and policy
15
Predefined first, custom only for a documented gap
16
Impersonation over keys, by default now
17
Keys are now blocked by default
18
Workforce vs. Workload Identity Federation
19
Hyperdisk is now the default, not Persistent Disk
20
OS Login and VM Manager: fleet control by IAM, not by hand
21
Spot VMs, custom machine types, and managed instance groups
22
GPUs, TPUs, and a snapshot schedule nobody had tested
23
Autopilot bills per Pod, Standard bills per node
24
Cloud Run: revisions, traffic splitting, and instant rollback
25
The Agent Runtime: the newest exam addition
26
Cloud Workstations and managed notebooks
27
Choosing between GKE, Cloud Run, and the Agent Runtime
28
Cloud Storage plus three file-storage products
29
Choosing a managed database, still the same tree
30
The relational family and the NoSQL split
31
Two newer additions: Managed Kafka and Memorystore
32
CMEK, a fleet-wide view, and a rotation that locked out a replica
33
Custom-mode VPCs, Shared VPC, and non-transitive peering
34
Cloud NGFW: Secure Tags replace network tags
35
Load balancer choice and an asymmetric tier rule
36
Cloud VPN, Interconnect, DNS, and NAT
37
Resizing subnets, static IPs, and static routes
38
Cloud Monitoring and Cloud Logging, still the backbone
39
Trace, Profiler, and Query Insights: a narrowing sequence
40
Personalized Service Health, Active Assist, and Cloud Hub
41
The traps that recur across every domain
42
Quick-fire recall
43
Exam-day strategy
44
Readiness checklist, fully current