feat(prover): Add support for scaling WGs and compressor

core/lib/config/src/configs/prover_autoscaler.rs

@@ -61,6 +61,8 @@ pub struct ProverAutoscalerScalerConfig {
     /// Duration after which pending pod considered long pending.
     #[serde(default = "ProverAutoscalerScalerConfig::default_long_pending_duration")]
     pub long_pending_duration: Duration,
+    /// List of simple autoscaler targets.
+    pub scaler_targets: Vec<ScalerTarget>,
 }
 
 #[derive(
@@ -93,6 +95,41 @@ pub enum Gpu {
     A100,
 }
 
+// TODO: generate this enum by QueueReport from https://github.com/matter-labs/zksync-era/blob/main/prover/crates/bin/prover_job_monitor/src/autoscaler_queue_reporter.rs#L23
+// and remove allowing of non_camel_case_types by generating field name parser.
+#[derive(Debug, Display, PartialEq, Eq, Hash, Clone, Deserialize, EnumString, Default)]
+#[allow(non_camel_case_types)]
+pub enum QueueReportFields {
+    #[strum(ascii_case_insensitive)]
+    basic_witness_jobs,
+    #[strum(ascii_case_insensitive)]
+    leaf_witness_jobs,
+    #[strum(ascii_case_insensitive)]
+    node_witness_jobs,
+    #[strum(ascii_case_insensitive)]
+    recursion_tip_witness_jobs,
+    #[strum(ascii_case_insensitive)]
+    scheduler_witness_jobs,
+    #[strum(ascii_case_insensitive)]
+    proof_compressor_jobs,
+    #[default]
+    #[strum(ascii_case_insensitive)]
+    prover_jobs,
+}
+
+/// ScalerTarget can be configured to autoscale any of services for which queue is reported by
+/// prover-job-monitor, except of provers. Provers need special treatment due to GPU requirement.
+#[derive(Debug, Clone, PartialEq, Deserialize, Default)]
+pub struct ScalerTarget {
+    pub queue_report_field: QueueReportFields,
+    pub pod_name_prefix: String,
+    /// Max replicas per cluster.
+    pub max_replicas: HashMap<String, usize>,
+    /// The queue will be divided by the speed and rounded up to get number of replicas.
+    #[serde(default = "ScalerTarget::default_speed")]
+    pub speed: usize,
+}
+
 impl ProverAutoscalerConfig {
     /// Default graceful shutdown timeout -- 5 seconds
     pub fn default_graceful_shutdown_timeout() -> Duration {
@@ -126,3 +163,9 @@ impl ProverAutoscalerScalerConfig {
         Duration::minutes(10)
     }
 }
+
+impl ScalerTarget {
+    pub fn default_speed() -> usize {
+        1
+    }
+}

core/lib/protobuf_config/src/proto/config/prover_autoscaler.proto

@@ -45,15 +45,28 @@ message MinProver {
   optional uint32 min = 2; // required
 }
 
+message MaxReplica {
+  optional string cluster = 1; // required
+  optional uint64 max = 2; // required
+}
+
+message ScalerTarget {
+  optional string queue_report_field = 1; // required
+  optional string pod_name_prefix = 2; // required
+  repeated MaxReplica max_replicas = 3; // required at least one
+  optional uint64 speed = 4; // optional
+}
+
 message ProverAutoscalerScalerConfig {
   optional uint32 prometheus_port = 1; // required
   optional std.Duration scaler_run_interval = 2; // optional
   optional string prover_job_monitor_url = 3; // required
   repeated string agents = 4; // required at least one
-  repeated ProtocolVersion protocol_versions = 5; // repeated at least one
+  repeated ProtocolVersion protocol_versions = 5; // required at least one
   repeated ClusterPriority cluster_priorities = 6; // optional
   repeated ProverSpeed prover_speed = 7; // optional
   optional uint32 long_pending_duration_s = 8; // optional
   repeated MaxProver max_provers = 9; // optional
   repeated MinProver min_provers = 10; // optional
+  repeated ScalerTarget scaler_targets = 11; // optional
 }

core/lib/protobuf_config/src/prover_autoscaler.rs

@@ -112,6 +112,12 @@ impl ProtoRepr for proto::ProverAutoscalerScalerConfig {
                 .map(|(i, e)| e.read().context(i))
                 .collect::<Result<_, _>>()
                 .context("min_provers")?,
+            scaler_targets: self
+                .scaler_targets
+                .iter()
+                .enumerate()
+                .map(|(i, x)| x.read().context(i).unwrap())
+                .collect::<Vec<_>>(),
         })
     }
 
@@ -151,6 +157,7 @@ impl ProtoRepr for proto::ProverAutoscalerScalerConfig {
                 .iter()
                 .map(|(k, v)| proto::MinProver::build(&(k.clone(), *v)))
                 .collect(),
+            scaler_targets: this.scaler_targets.iter().map(ProtoRepr::build).collect(),
         }
     }
 }
@@ -238,3 +245,57 @@ impl ProtoRepr for proto::MinProver {
         }
     }
 }
+
+impl ProtoRepr for proto::MaxReplica {
+    type Type = (String, usize);
+    fn read(&self) -> anyhow::Result<Self::Type> {
+        Ok((
+            required(&self.cluster).context("cluster")?.parse()?,
+            *required(&self.max).context("max")? as usize,
+        ))
+    }
+    fn build(this: &Self::Type) -> Self {
+        Self {
+            cluster: Some(this.0.to_string()),
+            max: Some(this.1 as u64),
+        }
+    }
+}
+
+impl ProtoRepr for proto::ScalerTarget {
+    type Type = configs::prover_autoscaler::ScalerTarget;
+    fn read(&self) -> anyhow::Result<Self::Type> {
+        Ok(Self::Type {
+            queue_report_field: required(&self.queue_report_field)
+                .and_then(|x| Ok((*x).parse()?))
+                .context("queue_report_field")?,
+            pod_name_prefix: required(&self.pod_name_prefix)
+                .context("pod_name_prefix")?
+                .clone(),
+            max_replicas: self
+                .max_replicas
+                .iter()
+                .enumerate()
+                .map(|(i, e)| e.read().context(i))
+                .collect::<Result<_, _>>()
+                .context("max_replicas")?,
+            speed: match self.speed {
+                Some(x) => x as usize,
+                None => Self::Type::default_speed(),
+            },
+        })
+    }
+
+    fn build(this: &Self::Type) -> Self {
+        Self {
+            queue_report_field: Some(this.queue_report_field.to_string()),
+            pod_name_prefix: Some(this.pod_name_prefix.clone()),
+            max_replicas: this
+                .max_replicas
+                .iter()
+                .map(|(k, v)| proto::MaxReplica::build(&(k.clone(), *v)))
+                .collect(),
+            speed: Some(this.speed as u64),
+        }
+    }
+}

prover/crates/bin/prover_autoscaler/src/cluster_types.rs

@@ -40,6 +40,7 @@ pub struct Namespace {
     #[serde(serialize_with = "ordered_map")]
     pub deployments: HashMap<String, Deployment>,
     pub pods: HashMap<String, Pod>,
+    #[serde(default)]
     pub scale_errors: Vec<ScaleEvent>,
 }
 
@@ -64,4 +65,5 @@ pub enum PodStatus {
     Pending,
     LongPending,
     NeedToMove,
+    Failed,
 }

prover/crates/bin/prover_autoscaler/src/global/queuer.rs

@@ -2,31 +2,48 @@ use std::collections::HashMap;
 
 use anyhow::{Context, Ok};
 use reqwest::Method;
-use zksync_prover_job_monitor::autoscaler_queue_reporter::VersionedQueueReport;
+use zksync_config::configs::prover_autoscaler::QueueReportFields;
+use zksync_prover_job_monitor::autoscaler_queue_reporter::{QueueReport, VersionedQueueReport};
+use zksync_types::prover_dal::JobCountStatistics;
 use zksync_utils::http_with_retries::send_request_with_retries;
 
 use crate::metrics::{AUTOSCALER_METRICS, DEFAULT_ERROR_CODE};
 
 const MAX_RETRIES: usize = 5;
 
-#[derive(Debug)]
-pub struct Queue {
-    pub queue: HashMap<String, u64>,
-}
+pub type Queue = HashMap<(String, QueueReportFields), u64>;
 
 #[derive(Default)]
 pub struct Queuer {
     pub prover_job_monitor_url: String,
 }
 
+fn sum_queue(jobs: JobCountStatistics) -> u64 {
+    (jobs.queued + jobs.in_progress) as u64
+}
+
+fn target_to_queue(target: QueueReportFields, report: &QueueReport) -> u64 {
+    match target {
+        QueueReportFields::basic_witness_jobs => sum_queue(report.basic_witness_jobs),
+        QueueReportFields::leaf_witness_jobs => sum_queue(report.leaf_witness_jobs),
+        QueueReportFields::node_witness_jobs => sum_queue(report.node_witness_jobs),
+        QueueReportFields::recursion_tip_witness_jobs => {
+            sum_queue(report.recursion_tip_witness_jobs)
+        }
+        QueueReportFields::scheduler_witness_jobs => sum_queue(report.scheduler_witness_jobs),
+        QueueReportFields::proof_compressor_jobs => sum_queue(report.proof_compressor_jobs),
+        QueueReportFields::prover_jobs => sum_queue(report.prover_jobs),
+    }
+}
+
 impl Queuer {
     pub fn new(pjm_url: String) -> Self {
         Self {
             prover_job_monitor_url: pjm_url,
         }
     }
 
-    pub async fn get_queue(&self) -> anyhow::Result<Queue> {
+    pub async fn get_queue(&self, jobs: &[QueueReportFields]) -> anyhow::Result<Queue> {
         let url = &self.prover_job_monitor_url;
         let response = send_request_with_retries(url, MAX_RETRIES, Method::GET, None, None).await;
         let response = response.map_err(|err| {
@@ -39,11 +56,16 @@ impl Queuer {
             .json::<Vec<VersionedQueueReport>>()
             .await
             .context("Failed to read response as json")?;
-        Ok(Queue {
-            queue: response
-                .iter()
-                .map(|x| (x.version.to_string(), x.report.prover_jobs.queued as u64))
-                .collect::<HashMap<_, _>>(),
-        })
+        Ok(response
+            .iter()
+            .flat_map(|x| {
+                jobs.iter().map(move |j| {
+                    (
+                        (x.version.to_string(), j.clone()),
+                        target_to_queue(j.clone(), &x.report),
+                    )
+                })
+            })
+            .collect::<HashMap<_, _>>())
     }
 }