feat(azure,gcp): give a cell's node pools a warm start

[IliaFeldgun]

Weekly e2e 33289951671 for the timings, every byoc cell in dd for the sizes. One row a cloud, for the default pool this module creates unless a program asks for more; every count is a pool total:

     pool                 start    ceiling  most  install job
                          was/now  was/now  used

aws -default-* 3 / 3 10 / 12 - fits 1800s
gcp pc-gke-np-default-* 2 / 3 20 / 12 9 1413s
azure default 1 / 3 10 / 12 6 1831s, exit 137

cold [n] --round-- [n n] --round-- [n n n] a VMSS or MIG create
and a cold image pull
for every round
warm [n n n] bought with the cluster

Most-used is the largest that pool has been on any cell in thirty days, from kubernetes_state.node.cpu_capacity per node. aws tags neither node group nor instance type, so its pool cannot be told from its cell, and its numbers are unchanged here anyway.

aws node groups are born at desired_size=3, so pinetools cluster install begins with capacity. AKS has no desired size and aks.py passed count=min_count, so azure's only pool started at a single node, and the GKE pool left initial_node_count unset. What the install job needed was Pending, and the autoscaler bought it in serial rounds - which is where azure spent the first twenty minutes of a 1800s deadline and was SIGKILLed, in both weekly runs. desired_size moves out of the AWS-only block, and both clouds read it through initial_count(), clamped to the pool's own min and max.

Three or four nodes is also the whole of it: each cell runs three more pools that pinetools creates through the operator - fdb, qroutersmol, indexbuilder

• so ours is three or four nodes of eleven. A 3-node seed is most of azure's gap to the four it had reached when the deadline killed it, and exactly the three gcp settles on.

GKE counted per zone, so min_size=1, max_size=10 was really a floor of two and a ceiling of twenty, and a seed would have had to be divided to land on three. It takes total limits instead - total_min_node_count and total_max_node_count with a location_policy, which GKE documents as totals where min_node_count is per zone - so every size is a pool total on all three clouds and initial_node_count takes initial_count() whole. Per zone was the house pattern when 46371fb wrote this module's GCP support; pinecone-io/iac moved to total limits in July, in 67fee389, and nothing carried it back.

The ceiling is 12 on every cloud now, three above the nine a gcp pool has actually used. gcp is the one that moves: it read as ten a zone and now means twelve, so a gcp pool above twelve would be scaled to it, and its floor drops from two nodes to one. aws and azure only gain two.

Neither seed is ours after the pool exists. AKS count and GKE initialNodeCount join ignore_changes, the way eks.py already ignores scalingConfig.desiredSize - and on GKE it matters more than drift, because a changed initialNodeCount replaces the node pool.

Reading one field in three places is what made this a three-cloud edit, so the NodePool to NodePoolConfig translation moves to common/node_pool.py, carrying over every field the two shapes share a name with. It also settles the taints: aws built a NodePoolTaint from any object with key, value and effect, while azure and gcp passed the caller's own objects into a pydantic list that takes only a dict or a NodePoolTaint, so tainting one of their pools raised a ValidationError before reaching a cloud. The public NodePool dataclasses stay per-cloud, each showing only the machine-size field its cloud reads, because a shared base would reorder positional args in a published API.

Problem

Describe the purpose of this change. What problem is being solved and why?

Solution

Describe the approach you took. Link to any relevant bugs, issues, docs, or other resources.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update
• Infrastructure change (CI configs, etc)
• Non-code change (docs, etc)
• None of the above: (explain here)

Test Plan

Describe specific steps for validating this change.

---

Note

Medium Risk
Changes default node counts and GKE autoscaling semantics (per-zone vs total caps), which can affect cluster capacity, install timing, and may scale existing GCP pools if they relied on higher per-zone ceilings; Pulumi ignore_changes reduces replacement risk for seed counts.

Overview
Warm start for Azure and GCP default pools so pinetools cluster install has nodes at creation instead of waiting on serial autoscaler rounds (which was timing out Azure e2e).

NodePoolConfig now treats desired_size as cross-cloud (default 3), adds initial_count() (clamped between min_size and max_size), and raises default max_size to 12. AKS agent pools use count=initial_count() instead of min_count; GKE sets initial_node_count and switches autoscaling to total_min_node_count / total_max_node_count so limits are pool totals across zones, not per zone.

Pulumi drift: AKS count and GKE initialNodeCount are added to ignore_changes (like EKS already does for desired size), since autoscaler owns live counts and GKE can replace pools if initialNodeCount changes.

Shared mapping: Per-cloud NodePool → NodePoolConfig duplication moves to common/node_pool.py (node_pool_configs), including consistent taint normalization (dict/object → NodePoolTaint). Azure/GCP public NodePool dataclasses gain desired_size; tests cover sizing and translation.

^{Reviewed by Cursor Bugbot for commit b1e02f9. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit b1e02f9. Configure here.}

[IliaFeldgun]

Azure still flaky, retriaging.