Feat: Add api to get machines with leaks

crates/api-db/migrations/20260316223033_machines_health_override_gin.sql

@@ -0,0 +1,2 @@
+-- Add a column to create a generalized inverted index on machines table for health_report_overrides
+CREATE INDEX IF NOT EXISTS machine_health_overrides_merges_gin_idx ON machines USING GIN ((health_report_overrides -> 'merges'));

crates/api-db/src/machine.rs

@@ -1657,6 +1657,14 @@ pub async fn find_machine_ids(
         qb.push(" INNER JOIN machine_topologies mt ON machines.id = mt.machine_id");
     }
 
+    // Return only machines that are powered on and have a health override with leak classification
+    if let Some(pstate) = &search_config.only_with_power_state {
+        let pstate_normalized = pstate.to_lowercase();
+        qb.push(" INNER JOIN power_options po ON po.host_id = machines.id AND po.last_fetched_power_state = ");
+        qb.push_bind(pstate_normalized);
+        qb.push("::host_power_state_t");
+    }
+
     qb.push(" WHERE TRUE");
 
     if search_config.only_maintenance {
@@ -1699,6 +1707,13 @@ pub async fn find_machine_ids(
         ));
     }
 
+    if let Some(ovrrd_str) = &search_config.only_with_health_alert {
+        qb.push(" AND health_report_overrides->'merges' ? ");
+        qb.push_bind(ovrrd_str.clone());
+        qb.push(" AND jsonb_array_length(health_report_overrides->'merges'->");
+        qb.push_bind(ovrrd_str);
+        qb.push("->'alerts') > 0");
+    }
     if search_config.mnnvl_only {
         qb.push(
             " AND mt.topology->'discovery_data'->'Info'->'dmi_data'->>'product_name' LIKE '%GB200%'",
@@ -1715,6 +1730,7 @@ pub async fn find_machine_ids(
     }
 
     let q = qb.build_query_as();
+
     let machine_ids: Vec<MachineId> = q
         .fetch_all(txn)
         .await

crates/api-model/src/machine/machine_search_config.rs

@@ -41,6 +41,8 @@ pub struct MachineSearchConfig {
     pub for_update: bool,
     // Only include NVLink capable machines (GB200/GB300 etc)
     pub mnnvl_only: bool,
+    pub only_with_power_state: Option<String>,
+    pub only_with_health_alert: Option<String>,
 }
 
 impl TryFrom<rpc::forge::MachineSearchConfig> for MachineSearchConfig {
@@ -63,6 +65,8 @@ impl TryFrom<rpc::forge::MachineSearchConfig> for MachineSearchConfig {
                 .transpose()?,
             for_update: false, // This isn't exposed to API callers
             mnnvl_only: value.mnnvl_only,
+            only_with_power_state: value.only_with_power_state,
+            only_with_health_alert: value.only_with_health_alert,
         })
     }
 }

crates/api/src/handlers/health.rs

@@ -44,16 +44,18 @@ pub async fn list_health_report_overrides(
 
     txn.commit().await?;
 
+    let ovr = host_machine
+        .health_report_overrides
+        .clone()
+        .into_iter()
+        .map(|o| HealthReportOverride {
+            report: Some(o.0.into()),
+            mode: o.1 as i32,
+        })
+        .collect();
+
     Ok(Response::new(rpc::ListHealthReportOverrideResponse {
-        overrides: host_machine
-            .health_report_overrides
-            .clone()
-            .into_iter()
-            .map(|o| HealthReportOverride {
-                report: Some(o.0.into()),
-                mode: o.1 as i32,
-            })
-            .collect(),
+        overrides: ovr,
     }))
 }
 
@@ -136,6 +138,7 @@ pub async fn insert_health_report_override(
         )
         .into());
     }
+
     let mut txn = api.txn_begin().await?;
 
     let mut report = health_report::HealthReport::try_from(report.clone())

crates/api/src/web/machine.rs

@@ -401,6 +401,8 @@ pub async fn fetch_machines(
         only_quarantine: false,
         instance_type_id: None,
         mnnvl_only: false,
+        only_with_power_state: None,
+        only_with_health_alert: None,
     });
 
     let machine_ids = api

crates/api/src/web/managed_host.rs

@@ -660,6 +660,8 @@ async fn fetch_managed_hosts_with_metadata(
             only_quarantine: false,
             instance_type_id: None,
             mnnvl_only: false,
+            only_with_health_alert: None,
+            only_with_power_state: None,
         }))
         .await?
         .into_inner()

crates/api/src/web/nvlink.rs

@@ -302,6 +302,8 @@ async fn fetch_logical_partitions(
                 only_quarantine: false,
                 instance_type_id: None,
                 mnnvl_only: true,
+                only_with_health_alert: None,
+                only_with_power_state: None,
             });
 
             let machine_ids = api

crates/health/Cargo.toml

@@ -69,6 +69,7 @@ logfmt = { path = "../logfmt" }
 carbide-version = { path = "../version" }
 
 [features]
+default = []
 bench-hooks = []
 
 [dev-dependencies]

crates/health/src/processor/health_report.rs

@@ -146,6 +146,7 @@ impl HealthReportProcessor {
                         calculated_status = ?state,
                         "Threshold check indicates issue but BMC reports sensor as OK - likely incorrect thresholds, reporting OK"
                     );
+
                     return SensorHealthResult::Success(HealthReportSuccess {
                         probe_id: Probe::Sensor,
                         target: Some(health.sensor_id.clone()),

crates/rpc/proto/forge.proto

@@ -2712,6 +2712,10 @@ message MachineSearchConfig {
   bool only_quarantine = 7;
   optional string instance_type_id = 8;
   bool mnnvl_only = 9;
+  // PowerState would mirror host_power_state_t, basically 'on' and 'off'
+  optional string only_with_power_state = 10;
+  // For example, to search for leak alerts, set the below to "hardware-health.tray-leak-detection"
+  optional string only_with_health_alert = 11;
 }
 
 message MachineStateHistoriesRequest {

crates/rpc/src/errors.rs

@@ -68,6 +68,8 @@ pub enum RpcDataConversionError {
     InvalidVpcVirtualizationType(i32),
     #[error("Invalid enum value received for critical error type: {0}")]
     InvalidCriticalErrorType(i32),
+    #[error("PowerState {0} is not valid")]
+    InvalidPowerState(i32),
     #[error("Instance ID {0} is not valid")]
     InvalidInstanceId(String),
     #[error("Remediation ID {0} is not valid")]