[compute_ctl] Improve 'empty' compute startup sequence (#4034)

Do several attempts to get spec from the control-plane and retry network
errors and all reasonable HTTP response codes. Do not hang waiting for
spec without confirmation from the control-plane that compute is known
and is in the `Empty` state.

Adjust the way we track `total_startup_ms` metric, it should be
calculated since the moment we received spec, not from the moment
`compute_ctl` started. Also introduce a new `wait_for_spec_ms` metric
to track the time spent sleeping and waiting for spec to be delivered
from control-plane.

Part of neondatabase/cloud#3533

compute_tools/src/bin/compute_ctl.rs

@@ -73,7 +73,7 @@ fn main() -> Result<()> {
     // Try to use just 'postgres' if no path is provided
     let pgbin = matches.get_one::<String>("pgbin").unwrap();
 
-    let mut spec = None;
+    let spec;
     let mut live_config_allowed = false;
     match spec_json {
         // First, try to get cluster spec from the cli argument
@@ -89,9 +89,13 @@ fn main() -> Result<()> {
             } else if let Some(id) = compute_id {
                 if let Some(cp_base) = control_plane_uri {
                     live_config_allowed = true;
-                    if let Ok(s) = get_spec_from_control_plane(cp_base, id) {
-                        spec = Some(s);
-                    }
+                    spec = match get_spec_from_control_plane(cp_base, id) {
+                        Ok(s) => s,
+                        Err(e) => {
+                            error!("cannot get response from control plane: {}", e);
+                            panic!("neither spec nor confirmation that compute is in the Empty state was received");
+                        }
+                    };
                 } else {
                     panic!("must specify both --control-plane-uri and --compute-id or none");
                 }
@@ -114,7 +118,6 @@ fn main() -> Result<()> {
         spec_set = false;
     }
     let compute_node = ComputeNode {
-        start_time: Utc::now(),
         connstr: Url::parse(connstr).context("cannot parse connstr as a URL")?,
         pgdata: pgdata.to_string(),
         pgbin: pgbin.to_string(),
@@ -147,6 +150,17 @@ fn main() -> Result<()> {
     let mut state = compute.state.lock().unwrap();
     let pspec = state.pspec.as_ref().expect("spec must be set");
     let startup_tracing_context = pspec.spec.startup_tracing_context.clone();
+
+    // Record for how long we slept waiting for the spec.
+    state.metrics.wait_for_spec_ms = Utc::now()
+        .signed_duration_since(state.start_time)
+        .to_std()
+        .unwrap()
+        .as_millis() as u64;
+    // Reset start time to the actual start of the configuration, so that
+    // total startup time was properly measured at the end.
+    state.start_time = Utc::now();
+
     state.status = ComputeStatus::Init;
     compute.state_changed.notify_all();
     drop(state);

compute_tools/src/compute.rs

@@ -38,7 +38,6 @@ use crate::spec::*;
 
 /// Compute node info shared across several `compute_ctl` threads.
 pub struct ComputeNode {
-    pub start_time: DateTime<Utc>,
     // Url type maintains proper escaping
     pub connstr: url::Url,
     pub pgdata: String,
@@ -66,6 +65,7 @@ pub struct ComputeNode {
 
 #[derive(Clone, Debug)]
 pub struct ComputeState {
+    pub start_time: DateTime<Utc>,
     pub status: ComputeStatus,
     /// Timestamp of the last Postgres activity
     pub last_active: DateTime<Utc>,
@@ -77,6 +77,7 @@ pub struct ComputeState {
 impl ComputeState {
     pub fn new() -> Self {
         Self {
+            start_time: Utc::now(),
             status: ComputeStatus::Empty,
             last_active: Utc::now(),
             error: None,
@@ -425,7 +426,7 @@ impl ComputeNode {
                 .unwrap()
                 .as_millis() as u64;
             state.metrics.total_startup_ms = startup_end_time
-                .signed_duration_since(self.start_time)
+                .signed_duration_since(compute_state.start_time)
                 .to_std()
                 .unwrap()
                 .as_millis() as u64;

compute_tools/src/http/api.rs

@@ -18,6 +18,7 @@ use tracing_utils::http::OtelName;
 
 fn status_response_from_state(state: &ComputeState) -> ComputeStatusResponse {
     ComputeStatusResponse {
+        start_time: state.start_time,
         tenant: state
             .pspec
             .as_ref()

compute_tools/src/http/openapi_spec.yaml

@@ -152,11 +152,14 @@ components:
       type: object
       description: Compute startup metrics.
       required:
+        - wait_for_spec_ms
         - sync_safekeepers_ms
         - basebackup_ms
         - config_ms
         - total_startup_ms
       properties:
+        wait_for_spec_ms:
+          type: integer
         sync_safekeepers_ms:
           type: integer
         basebackup_ms:
@@ -181,6 +184,13 @@ components:
         - status
         - last_active
       properties:
+        start_time:
+          type: string
+          description: |
+            Time when compute was started. If initially compute was started in the `empty`
+            state and then provided with valid spec, `start_time` will be reset to the
+            moment, when spec was received.
+          example: "2022-10-12T07:20:50.52Z"
         status:
           $ref: '#/components/schemas/ComputeStatus'
         last_active:

compute_tools/src/spec.rs

@@ -4,42 +4,117 @@ use std::str::FromStr;
 use anyhow::{anyhow, bail, Result};
 use postgres::config::Config;
 use postgres::{Client, NoTls};
-use tracing::{info, info_span, instrument, span_enabled, warn, Level};
+use reqwest::StatusCode;
+use tracing::{error, info, info_span, instrument, span_enabled, warn, Level};
 
 use crate::config;
 use crate::params::PG_HBA_ALL_MD5;
 use crate::pg_helpers::*;
 
-use compute_api::responses::ControlPlaneSpecResponse;
+use compute_api::responses::{ControlPlaneComputeStatus, ControlPlaneSpecResponse};
 use compute_api::spec::{ComputeSpec, Database, PgIdent, Role};
 
+// Do control plane request and return response if any. In case of error it
+// returns a bool flag indicating whether it makes sense to retry the request
+// and a string with error message.
+fn do_control_plane_request(
+    uri: &str,
+    jwt: &str,
+) -> Result<ControlPlaneSpecResponse, (bool, String)> {
+    let resp = reqwest::blocking::Client::new()
+        .get(uri)
+        .header("Authorization", jwt)
+        .send()
+        .map_err(|e| {
+            (
+                true,
+                format!("could not perform spec request to control plane: {}", e),
+            )
+        })?;
+
+    match resp.status() {
+        StatusCode::OK => match resp.json::<ControlPlaneSpecResponse>() {
+            Ok(spec_resp) => Ok(spec_resp),
+            Err(e) => Err((
+                true,
+                format!("could not deserialize control plane response: {}", e),
+            )),
+        },
+        StatusCode::SERVICE_UNAVAILABLE => {
+            Err((true, "control plane is temporarily unavailable".to_string()))
+        }
+        StatusCode::BAD_GATEWAY => {
+            // We have a problem with intermittent 502 errors now
+            // https://github.com/neondatabase/cloud/issues/2353
+            // It's fine to retry GET request in this case.
+            Err((true, "control plane request failed with 502".to_string()))
+        }
+        // Another code, likely 500 or 404, means that compute is unknown to the control plane
+        // or some internal failure happened. Doesn't make much sense to retry in this case.
+        _ => Err((
+            false,
+            format!(
+                "unexpected control plane response status code: {}",
+                resp.status()
+            ),
+        )),
+    }
+}
+
 /// Request spec from the control-plane by compute_id. If `NEON_CONSOLE_JWT`
 /// env variable is set, it will be used for authorization.
-pub fn get_spec_from_control_plane(base_uri: &str, compute_id: &str) -> Result<ComputeSpec> {
+pub fn get_spec_from_control_plane(
+    base_uri: &str,
+    compute_id: &str,
+) -> Result<Option<ComputeSpec>> {
     let cp_uri = format!("{base_uri}/management/api/v2/computes/{compute_id}/spec");
-    let jwt: String = match std::env::var("NEON_CONSOLE_JWT") {
+    let jwt: String = match std::env::var("NEON_CONTROL_PLANE_TOKEN") {
         Ok(v) => v,
         Err(_) => "".to_string(),
     };
+    let mut attempt = 1;
+    let mut spec: Result<Option<ComputeSpec>> = Ok(None);
+
     info!("getting spec from control plane: {}", cp_uri);
 
-    // TODO: check the response. We should distinguish cases when it's
-    // - network error, then retry
-    // - no spec for compute yet, then wait
-    // - compute id is unknown or any other error, then bail out
-    let resp: ControlPlaneSpecResponse = reqwest::blocking::Client::new()
-        .get(cp_uri)
-        .header("Authorization", jwt)
-        .send()
-        .map_err(|e| anyhow!("could not send spec request to control plane: {}", e))?
-        .json()
-        .map_err(|e| anyhow!("could not get compute spec from control plane: {}", e))?;
+    // Do 3 attempts to get spec from the control plane using the following logic:
+    // - network error -> then retry
+    // - compute id is unknown or any other error -> bail out
+    // - no spec for compute yet (Empty state) -> return Ok(None)
+    // - got spec -> return Ok(Some(spec))
+    while attempt < 4 {
+        spec = match do_control_plane_request(&cp_uri, &jwt) {
+            Ok(spec_resp) => match spec_resp.status {
+                ControlPlaneComputeStatus::Empty => Ok(None),
+                ControlPlaneComputeStatus::Attached => {
+                    if let Some(spec) = spec_resp.spec {
+                        Ok(Some(spec))
+                    } else {
+                        bail!("compute is attached, but spec is empty")
+                    }
+                }
+            },
+            Err((retry, msg)) => {
+                if retry {
+                    Err(anyhow!(msg))
+                } else {
+                    bail!(msg);
+                }
+            }
+        };
 
-    if let Some(spec) = resp.spec {
-        Ok(spec)
-    } else {
-        bail!("could not get compute spec from control plane")
+        if let Err(e) = &spec {
+            error!("attempt {} to get spec failed with: {}", attempt, e);
+        } else {
+            return spec;
+        }
+
+        attempt += 1;
+        std::thread::sleep(std::time::Duration::from_millis(100));
     }
+
+    // All attempts failed, return error.
+    spec
 }
 
 /// It takes cluster specification and does the following:

libs/compute_api/src/responses.rs

@@ -14,6 +14,7 @@ pub struct GenericAPIError {
 #[derive(Serialize, Debug)]
 #[serde(rename_all = "snake_case")]
 pub struct ComputeStatusResponse {
+    pub start_time: DateTime<Utc>,
     pub tenant: Option<String>,
     pub timeline: Option<String>,
     pub status: ComputeStatus,
@@ -63,6 +64,7 @@ where
 /// Response of the /metrics.json API
 #[derive(Clone, Debug, Default, Serialize)]
 pub struct ComputeMetrics {
+    pub wait_for_spec_ms: u64,
     pub sync_safekeepers_ms: u64,
     pub basebackup_ms: u64,
     pub config_ms: u64,
@@ -75,4 +77,16 @@ pub struct ComputeMetrics {
 #[derive(Deserialize, Debug)]
 pub struct ControlPlaneSpecResponse {
     pub spec: Option<ComputeSpec>,
+    pub status: ControlPlaneComputeStatus,
+}
+
+#[derive(Deserialize, Clone, Copy, Debug, PartialEq, Eq)]
+#[serde(rename_all = "snake_case")]
+pub enum ControlPlaneComputeStatus {
+    // Compute is known to control-plane, but it's not
+    // yet attached to any timeline / endpoint.
+    Empty,
+    // Compute is attached to some timeline / endpoint and
+    // should be able to start with provided spec.
+    Attached,
 }