feat(canopy): shm_size_floor spec + intent-driven floors

Re-land the shm-floor knob from the closed #80 on top of the CR
materialisation refactor. `compute_shm_and_shared_buffers` derives shm
from the pod's memory request/limit; for analytics workloads that
derived value is often lower than what shared_buffers benefits from,
and bumping memory request just to raise shm burns k8s scheduling
budget for no real gain.

Adds:
- `PostgresPhysicalReplicaSpec.shmSizeFloor: Option<Quantity>` — the
  Deployment builder now takes `max(computed_shm, floor)` and scales
  `shared_buffers_mb` proportionally so postgres actually uses the
  extra shm.
- `IntentConfig.shm_size_floor` (verify: 512Mi, analytics-dev / -dbt:
  2Gi) — materialised into the CR spec by
  IntentConfig::to_replica_spec.

Author: passcod

README.md

@@ -95,6 +95,7 @@ Defines a continuously-refreshed replica of a PostgreSQL database restored from
 | `storageSizeOverride` | `Quantity` | No | — | Override dynamic sizing with a fixed PVC size. When absent, PVC size is calculated from snapshot size. |
 | `storageSizeMaximum` | `Quantity` | No | `2Ti` | Maximum allowed PVC size. The restore will fail if the computed size exceeds this limit. |
 | `resources` | `ResourceRequirements` | No | — | CPU/memory resource requirements for the PostgreSQL pods. |
+| `shmSizeFloor` | `Quantity` | No | — | Floor on the postgres pod's `/dev/shm` sizing. When set, the Deployment uses `max(computed, shmSizeFloor)` — the computed value is derived from `resources` by [`compute_shm_and_shared_buffers`]. Useful when a workload's `shared_buffers` needs more shm than the resource-derived value provides, without wanting to bump the container's memory request. |
 | `serviceAnnotations` | `map[string]string` | No | — | Annotations applied to the Service. |
 | `podAnnotations` | `map[string]string` | No | — | Annotations applied to the PostgreSQL pods. |
 | `affinity` | `Affinity` | No | — | Pod scheduling affinity rules. |

crds.yaml

@@ -704,6 +704,17 @@ spec:
                   type: string
                 nullable: true
                 type: object
+              shmSizeFloor:
+                description: |-
+                  Floor on the postgres pod's `/dev/shm` sizing. When set, the
+                  Deployment builder uses `max(computed, shmSizeFloor)` — computed
+                  comes from [`compute_shm_and_shared_buffers`] driven by
+                  [`resources`]. Useful when the resource-derived value would be
+                  smaller than what a workload's `shared_buffers` needs (analytics
+                  / dbt) without wanting to bump the container's memory request
+                  upward just to raise shm.
+                nullable: true
+                x-kubernetes-int-or-string: true
               snapshotFilter:
                 nullable: true
                 properties:

src/controllers/canopy/intent.rs

@@ -41,6 +41,13 @@ pub struct IntentConfig {
 	pub service_annotations: Option<BTreeMap<String, String>>,
 	pub switchover_grace_period: TimeSpan,
 	pub storage_size_override: Quantity,
+	/// Floor on the postgres pod's `/dev/shm` sizing. Materialised into
+	/// `PostgresPhysicalReplicaSpec.shm_size_floor` so the shared
+	/// Deployment builder picks `max(computed_from_resources, floor)`.
+	/// Analytics workloads (`analytics-dev` / `analytics-dbt`) want a
+	/// higher shm than what a 2 GiB memory request would derive, without
+	/// paying the k8s scheduling cost of bumping the request.
+	pub shm_size_floor: Quantity,
 }
 
 fn resources(cpu_req: &str, mem_req: &str, cpu_lim: &str, mem_lim: &str) -> ResourceRequirements {
@@ -72,6 +79,7 @@ pub fn config_for(intent: &str) -> Option<IntentConfig> {
 			service_annotations: None,
 			switchover_grace_period: TimeSpan(Span::new().minutes(5)),
 			storage_size_override: Quantity("20Gi".to_string()),
+			shm_size_floor: Quantity("512Mi".to_string()),
 		}),
 		"analytics-dev" => Some(IntentConfig {
 			resources: Some(resources("500m", "2Gi", "4", "8Gi")),
@@ -81,6 +89,7 @@ pub fn config_for(intent: &str) -> Option<IntentConfig> {
 			service_annotations: None,
 			switchover_grace_period: TimeSpan(Span::new().minutes(5)),
 			storage_size_override: Quantity("50Gi".to_string()),
+			shm_size_floor: Quantity("2Gi".to_string()),
 		}),
 		"analytics-dbt" => Some(IntentConfig {
 			resources: Some(resources("500m", "2Gi", "4", "8Gi")),
@@ -96,6 +105,7 @@ pub fn config_for(intent: &str) -> Option<IntentConfig> {
 			])),
 			switchover_grace_period: TimeSpan(Span::new().minutes(2)),
 			storage_size_override: Quantity("50Gi".to_string()),
+			shm_size_floor: Quantity("2Gi".to_string()),
 		}),
 		_ => None,
 	}
@@ -141,6 +151,7 @@ impl IntentConfig {
 			storage_class: None,
 			storage_size_override: Some(self.storage_size_override.clone()),
 			resources: self.resources.clone(),
+			shm_size_floor: Some(self.shm_size_floor.clone()),
 			service_annotations: self
 				.service_annotations
 				.clone()

src/controllers/replica/scheduling.rs

@@ -267,6 +267,7 @@ mod tests {
 				storage_class: None,
 				storage_size_override: None,
 				resources: None,
+				shm_size_floor: None,
 				service_annotations: None,
 				pod_annotations: None,
 				affinity: None,

src/controllers/replica/schema_migration.rs

@@ -296,6 +296,7 @@ mod tests {
 				storage_class: None,
 				storage_size_override: None,
 				resources: None,
+				shm_size_floor: None,
 				service_annotations: None,
 				pod_annotations: None,
 				affinity: None,

src/controllers/replica/tests.rs

@@ -32,6 +32,7 @@ fn make_replica(
 			storage_class: None,
 			storage_size_override: None,
 			resources: None,
+			shm_size_floor: None,
 			service_annotations: None,
 			pod_annotations: None,
 			affinity: None,
@@ -191,6 +192,7 @@ fn snapshot_list_job_rotates_kopia_logs() {
 			storage_class: None,
 			storage_size_override: None,
 			resources: None,
+			shm_size_floor: None,
 			service_annotations: None,
 			pod_annotations: None,
 			affinity: None,

src/controllers/restore/builders.rs

@@ -952,7 +952,13 @@ pub fn build_deployment(
 	replica: &PostgresPhysicalReplica,
 ) -> Result<Deployment> {
 	let pvc_name = format!("{name}-data");
-	let (shm_size, shared_buffers_mb) = compute_shm_and_shared_buffers(&replica.spec.resources);
+	let (computed_shm, computed_shared_buffers_mb) =
+		compute_shm_and_shared_buffers(&replica.spec.resources);
+	let (shm_size, shared_buffers_mb) = apply_shm_floor(
+		&computed_shm,
+		computed_shared_buffers_mb,
+		replica.spec.shm_size_floor.as_ref(),
+	);
 	let creds_secret = SecretReference {
 		name: Some(format!("{}-creds", restore.spec.replica.name)),
 		namespace: Some(namespace.to_string()),
@@ -1001,6 +1007,35 @@ pub fn build_deployment(
 	})
 }
 
+/// Apply the caller's `shm_size_floor` (if any) to the resource-derived
+/// shm + shared_buffers pair. `shared_buffers` scales linearly with shm
+/// (postgres wants ~70 % of shm), so bumping shm bumps
+/// `shared_buffers_mb` proportionally.
+fn apply_shm_floor(
+	computed_shm: &Quantity,
+	computed_shared_buffers_mb: u64,
+	floor: Option<&Quantity>,
+) -> (Quantity, u64) {
+	let Some(floor) = floor else {
+		return (computed_shm.clone(), computed_shared_buffers_mb);
+	};
+	let computed_bytes = ParsedQuantity::try_from(computed_shm.clone())
+		.ok()
+		.and_then(|q| q.to_bytes_f64());
+	let floor_bytes = ParsedQuantity::try_from(floor.clone())
+		.ok()
+		.and_then(|q| q.to_bytes_f64());
+	match (computed_bytes, floor_bytes) {
+		(Some(c), Some(f)) if f > c => {
+			let ratio = f / c;
+			let scaled_shared_buffers =
+				((computed_shared_buffers_mb as f64) * ratio).floor() as u64;
+			(floor.clone(), scaled_shared_buffers.max(16))
+		}
+		_ => (computed_shm.clone(), computed_shared_buffers_mb),
+	}
+}
+
 /// Shared postgres Deployment builder — no CR dependencies. Both the
 /// CRD path (`build_deployment`) and the canopy path
 /// (`controllers::canopy::builders::build_canopy_postgres_deployment`) go

src/controllers/restore/tests.rs

@@ -31,6 +31,7 @@ fn deployment_uses_affinity_not_node_selector() {
 			storage_class: None,
 			storage_size_override: None,
 			resources: None,
+			shm_size_floor: None,
 			service_annotations: None,
 			pod_annotations: None,
 			affinity: None,
@@ -116,6 +117,7 @@ fn test_restore_and_replica() -> (PostgresPhysicalRestore, PostgresPhysicalRepli
 			storage_class: None,
 			storage_size_override: None,
 			resources: None,
+			shm_size_floor: None,
 			service_annotations: None,
 			pod_annotations: None,
 			affinity: None,

src/types/replica.rs

@@ -92,6 +92,16 @@ pub struct PostgresPhysicalReplicaSpec {
 	#[serde(default, skip_serializing_if = "Option::is_none")]
 	pub resources: Option<ResourceRequirements>,
 
+	/// Floor on the postgres pod's `/dev/shm` sizing. When set, the
+	/// Deployment builder uses `max(computed, shmSizeFloor)` — computed
+	/// comes from [`compute_shm_and_shared_buffers`] driven by
+	/// [`resources`]. Useful when the resource-derived value would be
+	/// smaller than what a workload's `shared_buffers` needs (analytics
+	/// / dbt) without wanting to bump the container's memory request
+	/// upward just to raise shm.
+	#[serde(default, skip_serializing_if = "Option::is_none")]
+	pub shm_size_floor: Option<Quantity>,
+
 	#[serde(default, skip_serializing_if = "Option::is_none")]
 	pub service_annotations: Option<BTreeMap<String, String>>,
 

tests/helpers.rs

@@ -152,6 +152,7 @@ pub fn build_replica(name: &str, secret_ref: &str, opts: ReplicaOpts) -> Postgre
 			storage_class: None,
 			storage_size_override: None,
 			resources: None,
+			shm_size_floor: None,
 			service_annotations: None,
 			pod_annotations: None,
 			affinity: None,