cleanup: no need for allow(dead_code) because we return the types

It was sensible in the LD_PRELOAD based impl, less so now.

.github/ansible/staging.eu-west-1.hosts.yaml

@@ -8,6 +8,14 @@ storage:
       pg_distrib_dir: /usr/local
       metric_collection_endpoint: http://neon-internal-api.aws.neon.build/billing/api/v1/usage_events
       metric_collection_interval: 10min
+      disk_usage_based_eviction:
+        max_usage_pct: 80
+        # TODO: learn typical resident-size growth rate [GiB/minute] and configure
+        # min_avail_bytes such that we have X minutes of headroom.
+        min_avail_bytes: 0
+        # We assume that the worst-case growth rate is small enough that we can
+        # catch above-threshold conditions by checking every 10s.
+        period: "10s"
       tenant_config:
         eviction_policy:
           kind: "LayerAccessThreshold"

.github/ansible/staging.us-east-2.hosts.yaml

@@ -8,6 +8,14 @@ storage:
       pg_distrib_dir: /usr/local
       metric_collection_endpoint: http://neon-internal-api.aws.neon.build/billing/api/v1/usage_events
       metric_collection_interval: 10min
+      disk_usage_based_eviction:
+        max_usage_pct: 80
+        # TODO: learn typical resident-size growth rate [GiB/minute] and configure
+        # min_avail_bytes such that we have X minutes of headroom.
+        min_avail_bytes: 0
+        # We assume that the worst-case growth rate is small enough that we can
+        # catch above-threshold conditions by checking every 10s.
+        period: "10s"
       tenant_config:
         eviction_policy:
           kind: "LayerAccessThreshold"

.github/workflows/build_and_test.yml

@@ -184,10 +184,10 @@ jobs:
           CARGO_FEATURES="--features testing"
           if [[ $BUILD_TYPE == "debug" ]]; then
             cov_prefix="scripts/coverage --profraw-prefix=$GITHUB_JOB --dir=/tmp/coverage run"
-            CARGO_FLAGS="--locked"
+            CARGO_FLAGS="--locked $CARGO_FEATURES"
           elif [[ $BUILD_TYPE == "release" ]]; then
             cov_prefix=""
-            CARGO_FLAGS="--locked --release"
+            CARGO_FLAGS="--locked --release $CARGO_FEATURES"
           fi
           echo "cov_prefix=${cov_prefix}" >> $GITHUB_ENV
           echo "CARGO_FEATURES=${CARGO_FEATURES}" >> $GITHUB_ENV
@@ -240,18 +240,11 @@ jobs:
 
       - name: Run cargo build
         run: |
-          ${cov_prefix} mold -run cargo build $CARGO_FLAGS $CARGO_FEATURES --bins --tests
+          ${cov_prefix} mold -run cargo build $CARGO_FLAGS --bins --tests
 
       - name: Run cargo test
         run: |
-          ${cov_prefix} cargo test $CARGO_FLAGS $CARGO_FEATURES
-
-          # Run separate tests for real S3
-          export ENABLE_REAL_S3_REMOTE_STORAGE=nonempty
-          export REMOTE_STORAGE_S3_BUCKET=neon-github-public-dev
-          export REMOTE_STORAGE_S3_REGION=eu-central-1
-          # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
-          ${cov_prefix} cargo test $CARGO_FLAGS --package remote_storage --test pagination_tests -- s3_pagination_should_work --exact
+          ${cov_prefix} cargo test $CARGO_FLAGS
 
       - name: Install rust binaries
         run: |
@@ -275,7 +268,7 @@ jobs:
             mkdir -p /tmp/neon/test_bin/
 
             test_exe_paths=$(
-              ${cov_prefix} cargo test $CARGO_FLAGS $CARGO_FEATURES --message-format=json --no-run |
+              ${cov_prefix} cargo test $CARGO_FLAGS --message-format=json --no-run |
               jq -r '.executable | select(. != null)'
             )
             for bin in $test_exe_paths; do

Cargo.lock

@@ -2474,6 +2474,7 @@ dependencies = [
  "strum",
  "strum_macros",
  "svg_fmt",
+ "sync_wrapper",
  "tempfile",
  "tenant_size_model",
  "thiserror",
@@ -3085,7 +3086,6 @@ dependencies = [
  "serde",
  "serde_json",
  "tempfile",
- "test-context",
  "tokio",
  "tokio-util",
  "toml_edit",
@@ -3889,27 +3889,6 @@ dependencies = [
  "winapi-util",
 ]
 
-[[package]]
-name = "test-context"
-version = "0.1.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "055831a02a4f5aa28fede67f2902014273eb8c21b958ac5ebbd59b71ef30dbc3"
-dependencies = [
- "async-trait",
- "futures",
- "test-context-macros",
-]
-
-[[package]]
-name = "test-context-macros"
-version = "0.1.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8901a55b0a7a06ebc4a674dcca925170da8e613fa3b163a1df804ed10afb154d"
-dependencies = [
- "quote",
- "syn",
-]
-
 [[package]]
 name = "textwrap"
 version = "0.16.0"
@@ -4556,6 +4535,7 @@ dependencies = [
  "once_cell",
  "pin-project-lite",
  "rand",
+ "regex",
  "routerify",
  "sentry",
  "serde",

Cargo.toml

@@ -97,7 +97,6 @@ strum_macros = "0.24"
 svg_fmt = "0.4.1"
 sync_wrapper = "0.1.2"
 tar = "0.4"
-test-context = "0.1"
 thiserror = "1.0"
 tls-listener = { version = "0.6", features = ["rustls", "hyper-h1"] }
 tokio = { version = "1.17", features = ["macros"] }

compute_tools/src/pg_helpers.rs

@@ -74,9 +74,18 @@ impl GenericOption {
     /// Represent `GenericOption` as configuration option.
     pub fn to_pg_setting(&self) -> String {
         if let Some(val) = &self.value {
+            // TODO: check in the console DB that we don't have these settings
+            // set for any non-deleted project and drop this override.
+            let name = match self.name.as_str() {
+                "safekeepers" => "neon.safekeepers",
+                "wal_acceptor_reconnect" => "neon.safekeeper_reconnect_timeout",
+                "wal_acceptor_connection_timeout" => "neon.safekeeper_connection_timeout",
+                it => it,
+            };
+
             match self.vartype.as_ref() {
-                "string" => format!("{} = '{}'", self.name, escape_conf_value(val)),
-                _ => format!("{} = {}", self.name, val),
+                "string" => format!("{} = '{}'", name, escape_conf_value(val)),
+                _ => format!("{} = {}", name, val),
             }
         } else {
             self.name.to_owned()

control_plane/src/compute.rs

@@ -403,7 +403,7 @@ impl PostgresNode {
 
     fn pg_ctl(&self, args: &[&str], auth_token: &Option<String>) -> Result<()> {
         let pg_ctl_path = self.env.pg_bin_dir(self.pg_version)?.join("pg_ctl");
-        let mut cmd = Command::new(&pg_ctl_path);
+        let mut cmd = Command::new(pg_ctl_path);
         cmd.args(
             [
                 &[
@@ -432,9 +432,7 @@ impl PostgresNode {
             cmd.env("NEON_AUTH_TOKEN", token);
         }
 
-        let pg_ctl = cmd
-            .output()
-            .context(format!("{} failed", pg_ctl_path.display()))?;
+        let pg_ctl = cmd.output().context("pg_ctl failed")?;
         if !pg_ctl.status.success() {
             anyhow::bail!(
                 "pg_ctl failed, exit code: {}, stdout: {}, stderr: {}",

control_plane/src/pageserver.rs

@@ -363,6 +363,11 @@ impl PageServerNode {
                 .map(|x| serde_json::from_str(x))
                 .transpose()
                 .context("Failed to parse 'eviction_policy' json")?,
+            min_resident_size_override: settings
+                .remove("min_resident_size_override")
+                .map(|x| x.parse::<u64>())
+                .transpose()
+                .context("Failed to parse 'min_resident_size_override' as integer")?,
         };
         if !settings.is_empty() {
             bail!("Unrecognized tenant settings: {settings:?}")
@@ -435,6 +440,11 @@ impl PageServerNode {
                     .map(|x| serde_json::from_str(x))
                     .transpose()
                     .context("Failed to parse 'eviction_policy' json")?,
+                min_resident_size_override: settings
+                    .get("min_resident_size_override")
+                    .map(|x| x.parse::<u64>())
+                    .transpose()
+                    .context("Failed to parse 'min_resident_size_override' as an integer")?,
             })
             .send()?
             .error_from_body()?;

control_plane/src/safekeeper.rs

@@ -156,7 +156,7 @@ impl SafekeeperNode {
         }
 
         background_process::start_process(
-            &format!("safekeeper-{id}"),
+            &format!("safekeeper {id}"),
             &datadir,
             &self.env.safekeeper_bin(),
             &args,

libs/pageserver_api/src/models.rs

@@ -120,6 +120,7 @@ pub struct TenantCreateRequest {
     // We might do that once the eviction feature has stabilizied.
     // For now, this field is not even documented in the openapi_spec.yml.
     pub eviction_policy: Option<serde_json::Value>,
+    pub min_resident_size_override: Option<u64>,
 }
 
 #[serde_as]
@@ -165,6 +166,7 @@ pub struct TenantConfigRequest {
     // We might do that once the eviction feature has stabilizied.
     // For now, this field is not even documented in the openapi_spec.yml.
     pub eviction_policy: Option<serde_json::Value>,
+    pub min_resident_size_override: Option<u64>,
 }
 
 impl TenantConfigRequest {
@@ -185,6 +187,7 @@ impl TenantConfigRequest {
             max_lsn_wal_lag: None,
             trace_read_requests: None,
             eviction_policy: None,
+            min_resident_size_override: None,
         }
     }
 }

libs/remote_storage/Cargo.toml

@@ -26,4 +26,3 @@ workspace_hack.workspace = true
 
 [dev-dependencies]
 tempfile.workspace = true
-test-context.workspace = true

libs/remote_storage/src/lib.rs

@@ -39,9 +39,6 @@ pub const DEFAULT_REMOTE_STORAGE_MAX_SYNC_ERRORS: u32 = 10;
 /// ~3500 PUT/COPY/POST/DELETE or 5500 GET/HEAD S3 requests
 /// https://aws.amazon.com/premiumsupport/knowledge-center/s3-request-limit-avoid-throttling/
 pub const DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT: usize = 100;
-/// No limits on the client side, which currenltly means 1000 for AWS S3.
-/// https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html#API_ListObjectsV2_RequestSyntax
-pub const DEFAULT_MAX_KEYS_PER_LIST_RESPONSE: Option<i32> = None;
 
 const REMOTE_STORAGE_PREFIX_SEPARATOR: char = '/';
 
@@ -67,10 +64,6 @@ impl RemotePath {
     pub fn object_name(&self) -> Option<&str> {
         self.0.file_name().and_then(|os_str| os_str.to_str())
     }
-
-    pub fn join(&self, segment: &Path) -> Self {
-        Self(self.0.join(segment))
-    }
 }
 
 /// Storage (potentially remote) API to manage its state.
@@ -273,7 +266,6 @@ pub struct S3Config {
     /// AWS S3 has various limits on its API calls, we need not to exceed those.
     /// See [`DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT`] for more details.
     pub concurrency_limit: NonZeroUsize,
-    pub max_keys_per_list_response: Option<i32>,
 }
 
 impl Debug for S3Config {
@@ -283,10 +275,6 @@ impl Debug for S3Config {
             .field("bucket_region", &self.bucket_region)
             .field("prefix_in_bucket", &self.prefix_in_bucket)
             .field("concurrency_limit", &self.concurrency_limit)
-            .field(
-                "max_keys_per_list_response",
-                &self.max_keys_per_list_response,
-            )
             .finish()
     }
 }
@@ -315,11 +303,6 @@ impl RemoteStorageConfig {
         )
         .context("Failed to parse 'concurrency_limit' as a positive integer")?;
 
-        let max_keys_per_list_response =
-            parse_optional_integer::<i32, _>("max_keys_per_list_response", toml)
-                .context("Failed to parse 'max_keys_per_list_response' as a positive integer")?
-                .or(DEFAULT_MAX_KEYS_PER_LIST_RESPONSE);
-
         let storage = match (local_path, bucket_name, bucket_region) {
             // no 'local_path' nor 'bucket_name' options are provided, consider this remote storage disabled
             (None, None, None) => return Ok(None),
@@ -341,7 +324,6 @@ impl RemoteStorageConfig {
                     .map(|endpoint| parse_toml_string("endpoint", endpoint))
                     .transpose()?,
                 concurrency_limit,
-                max_keys_per_list_response,
             }),
             (Some(local_path), None, None) => RemoteStorageKind::LocalFs(PathBuf::from(
                 parse_toml_string("local_path", local_path)?,

libs/remote_storage/src/s3_bucket.rs

@@ -102,7 +102,6 @@ pub struct S3Bucket {
     client: Client,
     bucket_name: String,
     prefix_in_bucket: Option<String>,
-    max_keys_per_list_response: Option<i32>,
     // Every request to S3 can be throttled or cancelled, if a certain number of requests per second is exceeded.
     // Same goes to IAM, which is queried before every S3 request, if enabled. IAM has even lower RPS threshold.
     // The helps to ensure we don't exceed the thresholds.
@@ -165,7 +164,6 @@ impl S3Bucket {
         Ok(Self {
             client,
             bucket_name: aws_config.bucket_name.clone(),
-            max_keys_per_list_response: aws_config.max_keys_per_list_response,
             prefix_in_bucket,
             concurrency_limiter: Arc::new(Semaphore::new(aws_config.concurrency_limit.get())),
         })
@@ -293,9 +291,7 @@ impl RemoteStorage for S3Bucket {
                 .list_objects_v2()
                 .bucket(self.bucket_name.clone())
                 .set_prefix(self.prefix_in_bucket.clone())
-                .delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string())
                 .set_continuation_token(continuation_token)
-                .set_max_keys(self.max_keys_per_list_response)
                 .send()
                 .await
                 .map_err(|e| {
@@ -310,7 +306,7 @@ impl RemoteStorage for S3Bucket {
                     .filter_map(|o| Some(self.s3_object_to_relative_path(o.key()?))),
             );
 
-            match fetch_response.next_continuation_token {
+            match fetch_response.continuation_token {
                 Some(new_token) => continuation_token = Some(new_token),
                 None => break,
             }
@@ -358,7 +354,6 @@ impl RemoteStorage for S3Bucket {
                 .set_prefix(list_prefix.clone())
                 .set_continuation_token(continuation_token)
                 .delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string())
-                .set_max_keys(self.max_keys_per_list_response)
                 .send()
                 .await
                 .map_err(|e| {
@@ -376,7 +371,7 @@ impl RemoteStorage for S3Bucket {
                     .filter_map(|o| Some(self.s3_object_to_relative_path(o.prefix()?))),
             );
 
-            match fetch_response.next_continuation_token {
+            match fetch_response.continuation_token {
                 Some(new_token) => continuation_token = Some(new_token),
                 None => break,
             }

libs/remote_storage/tests/pagination_tests.rs

@@ -1,275 +0,0 @@
-use std::collections::HashSet;
-use std::env;
-use std::num::{NonZeroU32, NonZeroUsize};
-use std::ops::ControlFlow;
-use std::path::{Path, PathBuf};
-use std::sync::Arc;
-use std::time::UNIX_EPOCH;
-
-use anyhow::Context;
-use remote_storage::{
-    GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind, S3Config,
-};
-use test_context::{test_context, AsyncTestContext};
-use tokio::task::JoinSet;
-use tracing::{debug, error, info};
-
-const ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME: &str = "ENABLE_REAL_S3_REMOTE_STORAGE";
-
-/// Tests that S3 client can list all prefixes, even if the response come paginated and requires multiple S3 queries.
-/// Uses real S3 and requires [`ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME`] and related S3 cred env vars specified.
-/// See the client creation in [`create_s3_client`] for details on the required env vars.
-/// If real S3 tests are disabled, the test passes, skipping any real test run: currently, there's no way to mark the test ignored in runtime with the
-/// deafult test framework, see https://github.com/rust-lang/rust/issues/68007 for details.
-///
-/// First, the test creates a set of S3 objects with keys `/${random_prefix_part}/${base_prefix_str}/sub_prefix_${i}/blob_${i}` in [`upload_s3_data`]
-/// where
-/// * `random_prefix_part` is set for the entire S3 client during the S3 client creation in [`create_s3_client`], to avoid multiple test runs interference
-/// * `base_prefix_str` is a common prefix to use in the client requests: we would want to ensure that the client is able to list nested prefixes inside the bucket
-///
-/// Then, verifies that the client does return correct prefixes when queried:
-/// * with no prefix, it lists everything after its `${random_prefix_part}/` — that should be `${base_prefix_str}` value only
-/// * with `${base_prefix_str}/` prefix, it lists every `sub_prefix_${i}`
-///
-/// With the real S3 enabled and `#[cfg(test)]` Rust configuration used, the S3 client test adds a `max-keys` param to limit the response keys.
-/// This way, we are able to test the pagination implicitly, by ensuring all results are returned from the remote storage and avoid uploading too many blobs to S3,
-/// since current default AWS S3 pagination limit is 1000.
-/// (see https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html#API_ListObjectsV2_RequestSyntax)
-///
-/// Lastly, the test attempts to clean up and remove all uploaded S3 files.
-/// If any errors appear during the clean up, they get logged, but the test is not failed or stopped until clean up is finished.
-#[test_context(MaybeEnabledS3)]
-#[tokio::test]
-async fn s3_pagination_should_work(ctx: &mut MaybeEnabledS3) -> anyhow::Result<()> {
-    let ctx = match ctx {
-        MaybeEnabledS3::Enabled(ctx) => ctx,
-        MaybeEnabledS3::Disabled => return Ok(()),
-        MaybeEnabledS3::UploadsFailed(e, _) => anyhow::bail!("S3 init failed: {e:?}"),
-    };
-
-    let test_client = Arc::clone(&ctx.client_with_excessive_pagination);
-    let expected_remote_prefixes = ctx.remote_prefixes.clone();
-
-    let base_prefix =
-        RemotePath::new(Path::new(ctx.base_prefix_str)).context("common_prefix construction")?;
-    let root_remote_prefixes = test_client
-        .list_prefixes(None)
-        .await
-        .context("client list root prefixes failure")?
-        .into_iter()
-        .collect::<HashSet<_>>();
-    assert_eq!(
-        root_remote_prefixes, HashSet::from([base_prefix.clone()]),
-        "remote storage root prefixes list mismatches with the uploads. Returned prefixes: {root_remote_prefixes:?}"
-    );
-
-    let nested_remote_prefixes = test_client
-        .list_prefixes(Some(&base_prefix))
-        .await
-        .context("client list nested prefixes failure")?
-        .into_iter()
-        .collect::<HashSet<_>>();
-    let remote_only_prefixes = nested_remote_prefixes
-        .difference(&expected_remote_prefixes)
-        .collect::<HashSet<_>>();
-    let missing_uploaded_prefixes = expected_remote_prefixes
-        .difference(&nested_remote_prefixes)
-        .collect::<HashSet<_>>();
-    assert_eq!(
-        remote_only_prefixes.len() + missing_uploaded_prefixes.len(), 0,
-        "remote storage nested prefixes list mismatches with the uploads. Remote only prefixes: {remote_only_prefixes:?}, missing uploaded prefixes: {missing_uploaded_prefixes:?}",
-    );
-
-    Ok(())
-}
-
-enum MaybeEnabledS3 {
-    Enabled(S3WithTestBlobs),
-    Disabled,
-    UploadsFailed(anyhow::Error, S3WithTestBlobs),
-}
-
-struct S3WithTestBlobs {
-    client_with_excessive_pagination: Arc<GenericRemoteStorage>,
-    base_prefix_str: &'static str,
-    remote_prefixes: HashSet<RemotePath>,
-    remote_blobs: HashSet<RemotePath>,
-}
-
-#[async_trait::async_trait]
-impl AsyncTestContext for MaybeEnabledS3 {
-    async fn setup() -> Self {
-        utils::logging::init(utils::logging::LogFormat::Test).expect("logging init failed");
-        if env::var(ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
-            info!(
-                "`{}` env variable is not set, skipping the test",
-                ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME
-            );
-            return Self::Disabled;
-        }
-
-        let max_keys_in_list_response = 10;
-        let upload_tasks_count = 1 + (2 * usize::try_from(max_keys_in_list_response).unwrap());
-
-        let client_with_excessive_pagination = create_s3_client(max_keys_in_list_response)
-            .context("S3 client creation")
-            .expect("S3 client creation failed");
-
-        let base_prefix_str = "test/";
-        match upload_s3_data(
-            &client_with_excessive_pagination,
-            base_prefix_str,
-            upload_tasks_count,
-        )
-        .await
-        {
-            ControlFlow::Continue(uploads) => {
-                info!("Remote objects created successfully");
-                Self::Enabled(S3WithTestBlobs {
-                    client_with_excessive_pagination,
-                    base_prefix_str,
-                    remote_prefixes: uploads.prefixes,
-                    remote_blobs: uploads.blobs,
-                })
-            }
-            ControlFlow::Break(uploads) => Self::UploadsFailed(
-                anyhow::anyhow!("One or multiple blobs failed to upload to S3"),
-                S3WithTestBlobs {
-                    client_with_excessive_pagination,
-                    base_prefix_str,
-                    remote_prefixes: uploads.prefixes,
-                    remote_blobs: uploads.blobs,
-                },
-            ),
-        }
-    }
-
-    async fn teardown(self) {
-        match self {
-            Self::Disabled => {}
-            Self::Enabled(ctx) | Self::UploadsFailed(_, ctx) => {
-                cleanup(&ctx.client_with_excessive_pagination, ctx.remote_blobs).await;
-            }
-        }
-    }
-}
-
-fn create_s3_client(max_keys_per_list_response: i32) -> anyhow::Result<Arc<GenericRemoteStorage>> {
-    let remote_storage_s3_bucket = env::var("REMOTE_STORAGE_S3_BUCKET")
-        .context("`REMOTE_STORAGE_S3_BUCKET` env var is not set, but real S3 tests are enabled")?;
-    let remote_storage_s3_region = env::var("REMOTE_STORAGE_S3_REGION")
-        .context("`REMOTE_STORAGE_S3_REGION` env var is not set, but real S3 tests are enabled")?;
-    let random_prefix_part = std::time::SystemTime::now()
-        .duration_since(UNIX_EPOCH)
-        .context("random s3 test prefix part calculation")?
-        .as_millis();
-    let remote_storage_config = RemoteStorageConfig {
-        max_concurrent_syncs: NonZeroUsize::new(100).unwrap(),
-        max_sync_errors: NonZeroU32::new(5).unwrap(),
-        storage: RemoteStorageKind::AwsS3(S3Config {
-            bucket_name: remote_storage_s3_bucket,
-            bucket_region: remote_storage_s3_region,
-            prefix_in_bucket: Some(format!("pagination_should_work_test_{random_prefix_part}/")),
-            endpoint: None,
-            concurrency_limit: NonZeroUsize::new(100).unwrap(),
-            max_keys_per_list_response: Some(max_keys_per_list_response),
-        }),
-    };
-    Ok(Arc::new(
-        GenericRemoteStorage::from_config(&remote_storage_config).context("remote storage init")?,
-    ))
-}
-
-struct Uploads {
-    prefixes: HashSet<RemotePath>,
-    blobs: HashSet<RemotePath>,
-}
-
-async fn upload_s3_data(
-    client: &Arc<GenericRemoteStorage>,
-    base_prefix_str: &'static str,
-    upload_tasks_count: usize,
-) -> ControlFlow<Uploads, Uploads> {
-    info!("Creating {upload_tasks_count} S3 files");
-    let mut upload_tasks = JoinSet::new();
-    for i in 1..upload_tasks_count + 1 {
-        let task_client = Arc::clone(client);
-        upload_tasks.spawn(async move {
-            let prefix = PathBuf::from(format!("{base_prefix_str}/sub_prefix_{i}/"));
-            let blob_prefix = RemotePath::new(&prefix)
-                .with_context(|| format!("{prefix:?} to RemotePath conversion"))?;
-            let blob_path = blob_prefix.join(Path::new(&format!("blob_{i}")));
-            debug!("Creating remote item {i} at path {blob_path:?}");
-
-            let data = format!("remote blob data {i}").into_bytes();
-            let data_len = data.len();
-            task_client
-                .upload(
-                    Box::new(std::io::Cursor::new(data)),
-                    data_len,
-                    &blob_path,
-                    None,
-                )
-                .await?;
-
-            Ok::<_, anyhow::Error>((blob_prefix, blob_path))
-        });
-    }
-
-    let mut upload_tasks_failed = false;
-    let mut uploaded_prefixes = HashSet::with_capacity(upload_tasks_count);
-    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
-    while let Some(task_run_result) = upload_tasks.join_next().await {
-        match task_run_result
-            .context("task join failed")
-            .and_then(|task_result| task_result.context("upload task failed"))
-        {
-            Ok((upload_prefix, upload_path)) => {
-                uploaded_prefixes.insert(upload_prefix);
-                uploaded_blobs.insert(upload_path);
-            }
-            Err(e) => {
-                error!("Upload task failed: {e:?}");
-                upload_tasks_failed = true;
-            }
-        }
-    }
-
-    let uploads = Uploads {
-        prefixes: uploaded_prefixes,
-        blobs: uploaded_blobs,
-    };
-    if upload_tasks_failed {
-        ControlFlow::Break(uploads)
-    } else {
-        ControlFlow::Continue(uploads)
-    }
-}
-
-async fn cleanup(client: &Arc<GenericRemoteStorage>, objects_to_delete: HashSet<RemotePath>) {
-    info!(
-        "Removing {} objects from the remote storage during cleanup",
-        objects_to_delete.len()
-    );
-    let mut delete_tasks = JoinSet::new();
-    for object_to_delete in objects_to_delete {
-        let task_client = Arc::clone(client);
-        delete_tasks.spawn(async move {
-            debug!("Deleting remote item at path {object_to_delete:?}");
-            task_client
-                .delete(&object_to_delete)
-                .await
-                .with_context(|| format!("{object_to_delete:?} removal"))
-        });
-    }
-
-    while let Some(task_run_result) = delete_tasks.join_next().await {
-        match task_run_result {
-            Ok(task_result) => match task_result {
-                Ok(()) => {}
-                Err(e) => error!("Delete task failed: {e:?}"),
-            },
-            Err(join_err) => error!("Delete task did not finish correctly: {join_err}"),
-        }
-    }
-}

libs/utils/Cargo.toml

@@ -19,6 +19,7 @@ jsonwebtoken.workspace = true
 nix.workspace = true
 once_cell.workspace = true
 pin-project-lite.workspace = true
+regex.workspace = true
 routerify.workspace = true
 serde.workspace = true
 serde_json.workspace = true

libs/utils/src/http/error.rs

@@ -20,9 +20,6 @@ pub enum ApiError {
     #[error("Conflict: {0}")]
     Conflict(String),
 
-    #[error("Precondition failed: {0}")]
-    PreconditionFailed(&'static str),
-
     #[error(transparent)]
     InternalServerError(anyhow::Error),
 }
@@ -47,10 +44,6 @@ impl ApiError {
             ApiError::Conflict(_) => {
                 HttpErrorBody::response_from_msg_and_status(self.to_string(), StatusCode::CONFLICT)
             }
-            ApiError::PreconditionFailed(_) => HttpErrorBody::response_from_msg_and_status(
-                self.to_string(),
-                StatusCode::PRECONDITION_FAILED,
-            ),
             ApiError::InternalServerError(err) => HttpErrorBody::response_from_msg_and_status(
                 err.to_string(),
                 StatusCode::INTERNAL_SERVER_ERROR,

libs/utils/src/lib.rs

@@ -51,6 +51,9 @@ pub mod history_buffer;
 
 pub mod measured_stream;
 
+pub mod serde_percent;
+pub mod serde_regex;
+
 /// use with fail::cfg("$name", "return(2000)")
 #[macro_export]
 macro_rules! failpoint_sleep_millis_async {

libs/utils/src/serde_percent.rs

@@ -0,0 +1,83 @@
+//! A serde::Deserialize type for percentages.
+//!
+//! See [`Percent`] for details.
+
+use serde::{Deserialize, Serialize};
+
+/// If the value is not an integer between 0 and 100,
+/// deserialization fails with a descriptive error.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash, Serialize, Deserialize)]
+#[serde(transparent)]
+pub struct Percent(#[serde(deserialize_with = "deserialize_pct_0_to_100")] u8);
+
+impl Percent {
+    pub fn get(&self) -> u8 {
+        self.0
+    }
+}
+
+fn deserialize_pct_0_to_100<'de, D>(deserializer: D) -> Result<u8, D::Error>
+where
+    D: serde::de::Deserializer<'de>,
+{
+    let v: u8 = serde::de::Deserialize::deserialize(deserializer)?;
+    if v > 100 {
+        return Err(serde::de::Error::custom(
+            "must be an integer between 0 and 100",
+        ));
+    }
+    Ok(v)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::Percent;
+
+    #[derive(serde::Deserialize, serde::Serialize, Debug, PartialEq, Eq)]
+    struct Foo {
+        bar: Percent,
+    }
+
+    #[test]
+    fn basics() {
+        let input = r#"{ "bar": 50 }"#;
+        let foo: Foo = serde_json::from_str(input).unwrap();
+        assert_eq!(foo.bar.get(), 50);
+    }
+    #[test]
+    fn null_handling() {
+        let input = r#"{ "bar": null }"#;
+        let res: Result<Foo, _> = serde_json::from_str(input);
+        assert!(res.is_err());
+    }
+    #[test]
+    fn zero() {
+        let input = r#"{ "bar": 0 }"#;
+        let foo: Foo = serde_json::from_str(input).unwrap();
+        assert_eq!(foo.bar.get(), 0);
+    }
+    #[test]
+    fn out_of_range_above() {
+        let input = r#"{ "bar": 101 }"#;
+        let res: Result<Foo, _> = serde_json::from_str(input);
+        assert!(res.is_err());
+    }
+    #[test]
+    fn out_of_range_below() {
+        let input = r#"{ "bar": -1 }"#;
+        let res: Result<Foo, _> = serde_json::from_str(input);
+        assert!(res.is_err());
+    }
+    #[test]
+    fn float() {
+        let input = r#"{ "bar": 50.5 }"#;
+        let res: Result<Foo, _> = serde_json::from_str(input);
+        assert!(res.is_err());
+    }
+    #[test]
+    fn string() {
+        let input = r#"{ "bar": "50 %" }"#;
+        let res: Result<Foo, _> = serde_json::from_str(input);
+        assert!(res.is_err());
+    }
+}

libs/utils/src/serde_regex.rs

@@ -0,0 +1,60 @@
+//! A `serde::{Deserialize,Serialize}` type for regexes.
+
+use std::ops::Deref;
+
+#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
+#[serde(transparent)]
+pub struct Regex(
+    #[serde(
+        deserialize_with = "deserialize_regex",
+        serialize_with = "serialize_regex"
+    )]
+    regex::Regex,
+);
+
+fn deserialize_regex<'de, D>(deserializer: D) -> Result<regex::Regex, D::Error>
+where
+    D: serde::de::Deserializer<'de>,
+{
+    let s: String = serde::de::Deserialize::deserialize(deserializer)?;
+    let re = regex::Regex::new(&s).map_err(serde::de::Error::custom)?;
+    Ok(re)
+}
+
+fn serialize_regex<S>(re: &regex::Regex, serializer: S) -> Result<S::Ok, S::Error>
+where
+    S: serde::ser::Serializer,
+{
+    serializer.collect_str(re.as_str())
+}
+
+impl Deref for Regex {
+    type Target = regex::Regex;
+
+    fn deref(&self) -> &regex::Regex {
+        &self.0
+    }
+}
+
+impl PartialEq for Regex {
+    fn eq(&self, other: &Regex) -> bool {
+        // comparing the automatons would be quite complicated
+        self.as_str() == other.as_str()
+    }
+}
+
+impl Eq for Regex {}
+
+#[cfg(test)]
+mod tests {
+
+    #[test]
+    fn roundtrip() {
+        let input = r#""foo.*bar""#;
+        let re: super::Regex = serde_json::from_str(input).unwrap();
+        assert!(re.is_match("foo123bar"));
+        assert!(!re.is_match("foo"));
+        let output = serde_json::to_string(&re).unwrap();
+        assert_eq!(output, input);
+    }
+}

libs/utils/src/signals.rs

@@ -1,7 +1,25 @@
+use signal_hook::flag;
 use signal_hook::iterator::Signals;
+use std::sync::atomic::AtomicBool;
+use std::sync::Arc;
 
 pub use signal_hook::consts::{signal::*, TERM_SIGNALS};
 
+pub fn install_shutdown_handlers() -> anyhow::Result<ShutdownSignals> {
+    let term_now = Arc::new(AtomicBool::new(false));
+    for sig in TERM_SIGNALS {
+        // When terminated by a second term signal, exit with exit code 1.
+        // This will do nothing the first time (because term_now is false).
+        flag::register_conditional_shutdown(*sig, 1, Arc::clone(&term_now))?;
+        // But this will "arm" the above for the second time, by setting it to true.
+        // The order of registering these is important, if you put this one first, it will
+        // first arm and then terminate ‒ all in the first round.
+        flag::register(*sig, Arc::clone(&term_now))?;
+    }
+
+    Ok(ShutdownSignals)
+}
+
 pub enum Signal {
     Quit,
     Interrupt,
@@ -21,7 +39,10 @@ impl Signal {
 pub struct ShutdownSignals;
 
 impl ShutdownSignals {
-    pub fn handle(mut handler: impl FnMut(Signal) -> anyhow::Result<()>) -> anyhow::Result<()> {
+    pub fn handle(
+        self,
+        mut handler: impl FnMut(Signal) -> anyhow::Result<()>,
+    ) -> anyhow::Result<()> {
         for raw_signal in Signals::new(TERM_SIGNALS)?.into_iter() {
             let signal = match raw_signal {
                 SIGINT => Signal::Interrupt,

pageserver/Cargo.toml

@@ -48,6 +48,7 @@ serde_json = { workspace = true, features = ["raw_value"] }
 serde_with.workspace = true
 signal-hook.workspace = true
 svg_fmt.workspace = true
+sync_wrapper.workspace = true
 tokio-tar.workspace = true
 thiserror.workspace = true
 tokio = { workspace = true, features = ["process", "sync", "fs", "rt", "io-util", "time"] }

pageserver/src/bin/pageserver.rs

@@ -8,6 +8,7 @@ use anyhow::{anyhow, Context};
 use clap::{Arg, ArgAction, Command};
 use fail::FailScenario;
 use metrics::launch_timestamp::{set_launch_timestamp_metric, LaunchTimestamp};
+use pageserver::disk_usage_eviction_task::{self, launch_disk_usage_global_eviction_task};
 use remote_storage::GenericRemoteStorage;
 use tracing::*;
 
@@ -24,9 +25,11 @@ use pageserver::{
     virtual_file,
 };
 use postgres_backend::AuthType;
-use utils::signals::ShutdownSignals;
 use utils::{
-    auth::JwtAuth, logging, project_git_version, sentry_init::init_sentry, signals::Signal,
+    auth::JwtAuth,
+    logging, project_git_version,
+    sentry_init::init_sentry,
+    signals::{self, Signal},
     tcp_listener,
 };
 
@@ -261,6 +264,9 @@ fn start_pageserver(
     info!("Starting pageserver pg protocol handler on {pg_addr}");
     let pageserver_listener = tcp_listener::bind(pg_addr)?;
 
+    // Install signal handlers
+    let signals = signals::install_shutdown_handlers()?;
+
     // Launch broker client
     WALRECEIVER_RUNTIME.block_on(pageserver::broker_client::init_broker_client(conf))?;
 
@@ -314,14 +320,34 @@ fn start_pageserver(
     // Scan the local 'tenants/' directory and start loading the tenants
     BACKGROUND_RUNTIME.block_on(mgr::init_tenant_mgr(conf, remote_storage.clone()))?;
 
+    // shared state between the disk-usage backed eviction background task and the http endpoint
+    // that allows triggering disk-usage based eviction manually. note that the http endpoint
+    // is still accessible even if background task is not configured as long as remote storage has
+    // been configured.
+    let disk_usage_eviction_state: Arc<disk_usage_eviction_task::State> = Arc::default();
+
+    if let Some(remote_storage) = &remote_storage {
+        launch_disk_usage_global_eviction_task(
+            conf,
+            remote_storage.clone(),
+            disk_usage_eviction_state.clone(),
+        )?;
+    }
+
     // Start up the service to handle HTTP mgmt API request. We created the
     // listener earlier already.
     {
         let _rt_guard = MGMT_REQUEST_RUNTIME.enter();
 
-        let router = http::make_router(conf, launch_ts, http_auth, remote_storage)?
-            .build()
-            .map_err(|err| anyhow!(err))?;
+        let router = http::make_router(
+            conf,
+            launch_ts,
+            http_auth,
+            remote_storage,
+            disk_usage_eviction_state,
+        )?
+        .build()
+        .map_err(|err| anyhow!(err))?;
         let service = utils::http::RouterService::new(router).unwrap();
         let server = hyper::Server::from_tcp(http_listener)?
             .serve(service)
@@ -404,7 +430,7 @@ fn start_pageserver(
     }
 
     // All started up! Now just sit and wait for shutdown signal.
-    ShutdownSignals::handle(|signal| match signal {
+    signals.handle(|signal| match signal {
         Signal::Quit => {
             info!(
                 "Got {}. Terminating in immediate shutdown mode",

pageserver/src/config.rs

@@ -27,6 +27,7 @@ use utils::{
     logging::LogFormat,
 };
 
+use crate::disk_usage_eviction_task::DiskUsageEvictionTaskConfig;
 use crate::tenant::config::TenantConf;
 use crate::tenant::config::TenantConfOpt;
 use crate::tenant::{TENANT_ATTACHING_MARKER_FILENAME, TIMELINES_SEGMENT_NAME};
@@ -92,6 +93,8 @@ pub mod defaults {
 
 #evictions_low_residence_duration_metric_threshold = '{DEFAULT_EVICTIONS_LOW_RESIDENCE_DURATION_METRIC_THRESHOLD}'
 
+#disk_usage_based_eviction = {{ max_usage_pct = .., min_avail_bytes = .., period = "10s"}}
+
 # [tenant_config]
 #checkpoint_distance = {DEFAULT_CHECKPOINT_DISTANCE} # in bytes
 #checkpoint_timeout = {DEFAULT_CHECKPOINT_TIMEOUT}
@@ -104,6 +107,8 @@ pub mod defaults {
 #image_creation_threshold = {DEFAULT_IMAGE_CREATION_THRESHOLD}
 #pitr_interval = '{DEFAULT_PITR_INTERVAL}'
 
+#min_resident_size_override = .. # in bytes
+
 # [remote_storage]
 
 "###
@@ -165,10 +170,6 @@ pub struct PageServerConf {
 
     /// Number of concurrent [`Tenant::gather_size_inputs`] allowed.
     pub concurrent_tenant_size_logical_size_queries: ConfigurableSemaphore,
-    /// Limit of concurrent [`Tenant::gather_size_inputs`] issued by module `eviction_task`.
-    /// The number of permits is the same as `concurrent_tenant_size_logical_size_queries`.
-    /// See the comment in `eviction_task` for details.
-    pub eviction_task_immitated_concurrent_logical_size_queries: ConfigurableSemaphore,
 
     // How often to collect metrics and send them to the metrics endpoint.
     pub metric_collection_interval: Duration,
@@ -180,6 +181,8 @@ pub struct PageServerConf {
     // See the corresponding metric's help string.
     pub evictions_low_residence_duration_metric_threshold: Duration,
 
+    pub disk_usage_based_eviction: Option<DiskUsageEvictionTaskConfig>,
+
     pub test_remote_failures: u64,
 
     pub ondemand_download_behavior_treat_error_as_warn: bool,
@@ -243,7 +246,7 @@ struct PageServerConfigBuilder {
 
     log_format: BuilderValue<LogFormat>,
 
-    concurrent_tenant_size_logical_size_queries: BuilderValue<NonZeroUsize>,
+    concurrent_tenant_size_logical_size_queries: BuilderValue<ConfigurableSemaphore>,
 
     metric_collection_interval: BuilderValue<Duration>,
     cached_metric_collection_interval: BuilderValue<Duration>,
@@ -252,6 +255,8 @@ struct PageServerConfigBuilder {
 
     evictions_low_residence_duration_metric_threshold: BuilderValue<Duration>,
 
+    disk_usage_based_eviction: BuilderValue<Option<DiskUsageEvictionTaskConfig>>,
+
     test_remote_failures: BuilderValue<u64>,
 
     ondemand_download_behavior_treat_error_as_warn: BuilderValue<bool>,
@@ -290,9 +295,7 @@ impl Default for PageServerConfigBuilder {
             .expect("cannot parse default keepalive interval")),
             log_format: Set(LogFormat::from_str(DEFAULT_LOG_FORMAT).unwrap()),
 
-            concurrent_tenant_size_logical_size_queries: Set(
-                ConfigurableSemaphore::DEFAULT_INITIAL,
-            ),
+            concurrent_tenant_size_logical_size_queries: Set(ConfigurableSemaphore::default()),
             metric_collection_interval: Set(humantime::parse_duration(
                 DEFAULT_METRIC_COLLECTION_INTERVAL,
             )
@@ -312,6 +315,8 @@ impl Default for PageServerConfigBuilder {
             )
             .expect("cannot parse DEFAULT_EVICTIONS_LOW_RESIDENCE_DURATION_METRIC_THRESHOLD")),
 
+            disk_usage_based_eviction: Set(None),
+
             test_remote_failures: Set(0),
 
             ondemand_download_behavior_treat_error_as_warn: Set(false),
@@ -395,7 +400,7 @@ impl PageServerConfigBuilder {
         self.log_format = BuilderValue::Set(log_format)
     }
 
-    pub fn concurrent_tenant_size_logical_size_queries(&mut self, u: NonZeroUsize) {
+    pub fn concurrent_tenant_size_logical_size_queries(&mut self, u: ConfigurableSemaphore) {
         self.concurrent_tenant_size_logical_size_queries = BuilderValue::Set(u);
     }
 
@@ -431,6 +436,10 @@ impl PageServerConfigBuilder {
         self.evictions_low_residence_duration_metric_threshold = BuilderValue::Set(value);
     }
 
+    pub fn disk_usage_based_eviction(&mut self, value: Option<DiskUsageEvictionTaskConfig>) {
+        self.disk_usage_based_eviction = BuilderValue::Set(value);
+    }
+
     pub fn ondemand_download_behavior_treat_error_as_warn(
         &mut self,
         ondemand_download_behavior_treat_error_as_warn: bool,
@@ -440,11 +449,6 @@ impl PageServerConfigBuilder {
     }
 
     pub fn build(self) -> anyhow::Result<PageServerConf> {
-        let concurrent_tenant_size_logical_size_queries = self
-            .concurrent_tenant_size_logical_size_queries
-            .ok_or(anyhow!(
-                "missing concurrent_tenant_size_logical_size_queries"
-            ))?;
         Ok(PageServerConf {
             listen_pg_addr: self
                 .listen_pg_addr
@@ -492,12 +496,11 @@ impl PageServerConfigBuilder {
                 .broker_keepalive_interval
                 .ok_or(anyhow!("No broker keepalive interval provided"))?,
             log_format: self.log_format.ok_or(anyhow!("missing log_format"))?,
-            concurrent_tenant_size_logical_size_queries: ConfigurableSemaphore::new(
-                concurrent_tenant_size_logical_size_queries,
-            ),
-            eviction_task_immitated_concurrent_logical_size_queries: ConfigurableSemaphore::new(
-                concurrent_tenant_size_logical_size_queries,
-            ),
+            concurrent_tenant_size_logical_size_queries: self
+                .concurrent_tenant_size_logical_size_queries
+                .ok_or(anyhow!(
+                    "missing concurrent_tenant_size_logical_size_queries"
+                ))?,
             metric_collection_interval: self
                 .metric_collection_interval
                 .ok_or(anyhow!("missing metric_collection_interval"))?,
@@ -515,6 +518,9 @@ impl PageServerConfigBuilder {
                 .ok_or(anyhow!(
                     "missing evictions_low_residence_duration_metric_threshold"
                 ))?,
+            disk_usage_based_eviction: self
+                .disk_usage_based_eviction
+                .ok_or(anyhow!("missing disk_usage_based_eviction"))?,
             test_remote_failures: self
                 .test_remote_failures
                 .ok_or(anyhow!("missing test_remote_failuers"))?,
@@ -692,7 +698,8 @@ impl PageServerConf {
                 "concurrent_tenant_size_logical_size_queries" => builder.concurrent_tenant_size_logical_size_queries({
                     let input = parse_toml_string(key, item)?;
                     let permits = input.parse::<usize>().context("expected a number of initial permits, not {s:?}")?;
-                    NonZeroUsize::new(permits).context("initial semaphore permits out of range: 0, use other configuration to disable a feature")?
+                    let permits = NonZeroUsize::new(permits).context("initial semaphore permits out of range: 0, use other configuration to disable a feature")?;
+                    ConfigurableSemaphore::new(permits)
                 }),
                 "metric_collection_interval" => builder.metric_collection_interval(parse_toml_duration(key, item)?),
                 "cached_metric_collection_interval" => builder.cached_metric_collection_interval(parse_toml_duration(key, item)?),
@@ -704,6 +711,12 @@ impl PageServerConf {
                     builder.synthetic_size_calculation_interval(parse_toml_duration(key, item)?),
                 "test_remote_failures" => builder.test_remote_failures(parse_toml_u64(key, item)?),
                 "evictions_low_residence_duration_metric_threshold" => builder.evictions_low_residence_duration_metric_threshold(parse_toml_duration(key, item)?),
+                "disk_usage_based_eviction" => {
+                    tracing::info!("disk_usage_based_eviction: {:#?}", &item);
+                    builder.disk_usage_based_eviction(
+                    toml_edit::de::from_item(item.clone())
+                    .context("parse disk_usage_based_eviction")?)
+                },
                 "ondemand_download_behavior_treat_error_as_warn" => builder.ondemand_download_behavior_treat_error_as_warn(parse_toml_bool(key, item)?),
                 _ => bail!("unrecognized pageserver option '{key}'"),
             }
@@ -808,6 +821,13 @@ impl PageServerConf {
             );
         }
 
+        if let Some(item) = item.get("min_resident_size_override") {
+            t_conf.min_resident_size_override = Some(
+                toml_edit::de::from_item(item.clone())
+                    .context("parse min_resident_size_override")?,
+            );
+        }
+
         Ok(t_conf)
     }
 
@@ -840,8 +860,6 @@ impl PageServerConf {
             broker_keepalive_interval: Duration::from_secs(5000),
             log_format: LogFormat::from_str(defaults::DEFAULT_LOG_FORMAT).unwrap(),
             concurrent_tenant_size_logical_size_queries: ConfigurableSemaphore::default(),
-            eviction_task_immitated_concurrent_logical_size_queries: ConfigurableSemaphore::default(
-            ),
             metric_collection_interval: Duration::from_secs(60),
             cached_metric_collection_interval: Duration::from_secs(60 * 60),
             metric_collection_endpoint: defaults::DEFAULT_METRIC_COLLECTION_ENDPOINT,
@@ -850,6 +868,7 @@ impl PageServerConf {
                 defaults::DEFAULT_EVICTIONS_LOW_RESIDENCE_DURATION_METRIC_THRESHOLD,
             )
             .unwrap(),
+            disk_usage_based_eviction: None,
             test_remote_failures: 0,
             ondemand_download_behavior_treat_error_as_warn: false,
         }
@@ -934,11 +953,6 @@ impl ConfigurableSemaphore {
             inner: std::sync::Arc::new(tokio::sync::Semaphore::new(initial_permits.get())),
         }
     }
-
-    /// Returns the configured amount of permits.
-    pub fn initial_permits(&self) -> NonZeroUsize {
-        self.initial_permits
-    }
 }
 
 impl Default for ConfigurableSemaphore {
@@ -1043,8 +1057,6 @@ log_format = 'json'
                 )?,
                 log_format: LogFormat::from_str(defaults::DEFAULT_LOG_FORMAT).unwrap(),
                 concurrent_tenant_size_logical_size_queries: ConfigurableSemaphore::default(),
-                eviction_task_immitated_concurrent_logical_size_queries:
-                    ConfigurableSemaphore::default(),
                 metric_collection_interval: humantime::parse_duration(
                     defaults::DEFAULT_METRIC_COLLECTION_INTERVAL
                 )?,
@@ -1058,6 +1070,7 @@ log_format = 'json'
                 evictions_low_residence_duration_metric_threshold: humantime::parse_duration(
                     defaults::DEFAULT_EVICTIONS_LOW_RESIDENCE_DURATION_METRIC_THRESHOLD
                 )?,
+                disk_usage_based_eviction: None,
                 test_remote_failures: 0,
                 ondemand_download_behavior_treat_error_as_warn: false,
             },
@@ -1105,13 +1118,12 @@ log_format = 'json'
                 broker_keepalive_interval: Duration::from_secs(5),
                 log_format: LogFormat::Json,
                 concurrent_tenant_size_logical_size_queries: ConfigurableSemaphore::default(),
-                eviction_task_immitated_concurrent_logical_size_queries:
-                    ConfigurableSemaphore::default(),
                 metric_collection_interval: Duration::from_secs(222),
                 cached_metric_collection_interval: Duration::from_secs(22200),
                 metric_collection_endpoint: Some(Url::parse("http://localhost:80/metrics")?),
                 synthetic_size_calculation_interval: Duration::from_secs(333),
                 evictions_low_residence_duration_metric_threshold: Duration::from_secs(444),
+                disk_usage_based_eviction: None,
                 test_remote_failures: 0,
                 ondemand_download_behavior_treat_error_as_warn: false,
             },
@@ -1238,7 +1250,6 @@ broker_endpoint = '{broker_endpoint}'
                         prefix_in_bucket: Some(prefix_in_bucket.clone()),
                         endpoint: Some(endpoint.clone()),
                         concurrency_limit: s3_concurrency_limit,
-                        max_keys_per_list_response: None,
                     }),
                 },
                 "Remote storage config should correctly parse the S3 config"

pageserver/src/disk_usage_eviction_task.rs

@@ -0,0 +1,679 @@
+//! This module implements the pageserver-global disk-usage-based layer eviction task.
+//!
+//! # Mechanics
+//!
+//! Function `launch_disk_usage_global_eviction_task` starts a pageserver-global background
+//! loop that evicts layers in response to a shortage of available bytes
+//! in the $repo/tenants directory's filesystem.
+//!
+//! The loop runs periodically at a configurable `period`.
+//!
+//! Each loop iteration uses `statvfs` to determine filesystem-level space usage.
+//! It compares the returned usage data against two different types of thresholds.
+//! The iteration tries to evict layers until app-internal accounting says we should be below the thresholds.
+//! We cross-check this internal accounting with the real world by making another `statvfs` at the end of the iteration.
+//! We're good if that second statvfs shows that we're _actually_ below the configured thresholds.
+//! If we're still above one or more thresholds, we emit a warning log message, leaving it to the operator to investigate further.
+//!
+//! # Eviction Policy
+//!
+//! There are two thresholds:
+//! `max_usage_pct` is the relative available space, expressed in percent of the total filesystem space.
+//! If the actual usage is higher, the threshold is exceeded.
+//! `min_avail_bytes` is the absolute available space in bytes.
+//! If the actual usage is lower, the threshold is exceeded.
+//!
+//! The iteration evicts layers in LRU fashion, but, with a weak reservation per tenant.
+//! The reservation is to keep the most recently accessed X bytes per tenant resident.
+//! If we cannot relieve pressure by evicting layers outside of the reservation, we
+//! start evicting layers that are part of the reservation, LRU first.
+//!
+//! The value for the per-tenant reservation is referred to as `tenant_min_resident_size`
+//! throughout the code, but, no actual variable carries that name.
+//! The per-tenant default value is the `max(tenant's layer file sizes, regardless of local or remote)`.
+//! The idea is to allow at least one layer to be resident per tenant, to ensure it can make forward progress
+//! during page reconstruction.
+//! An alternative default for all tenants can be specified in the `tenant_config` section of the config.
+//! Lastly, each tenant can have an override in their respectice tenant config (`min_resident_size_override`).
+
+use std::{
+    collections::HashMap,
+    path::Path,
+    sync::Arc,
+    time::{Duration, SystemTime},
+};
+
+use anyhow::Context;
+use remote_storage::GenericRemoteStorage;
+use serde::{Deserialize, Serialize};
+use tokio::time::Instant;
+use tokio_util::sync::CancellationToken;
+use tracing::{debug, error, info, instrument, warn, Instrument};
+use utils::serde_percent::Percent;
+
+use crate::{
+    config::PageServerConf,
+    task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
+    tenant::{self, storage_layer::PersistentLayer, Timeline},
+};
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct DiskUsageEvictionTaskConfig {
+    pub max_usage_pct: Percent,
+    pub min_avail_bytes: u64,
+    #[serde(with = "humantime_serde")]
+    pub period: Duration,
+    #[cfg(feature = "testing")]
+    pub mock_statvfs: Option<crate::statvfs::mock::Behavior>,
+}
+
+#[derive(Default)]
+pub struct State {
+    /// Exclude http requests and background task from running at the same time.
+    mutex: tokio::sync::Mutex<()>,
+}
+
+pub fn launch_disk_usage_global_eviction_task(
+    conf: &'static PageServerConf,
+    storage: GenericRemoteStorage,
+    state: Arc<State>,
+) -> anyhow::Result<()> {
+    let Some(task_config) = &conf.disk_usage_based_eviction else {
+        info!("disk usage based eviction task not configured");
+        return Ok(());
+    };
+
+    info!("launching disk usage based eviction task");
+
+    task_mgr::spawn(
+        BACKGROUND_RUNTIME.handle(),
+        TaskKind::DiskUsageEviction,
+        None,
+        None,
+        "disk usage based eviction",
+        false,
+        async move {
+            disk_usage_eviction_task(
+                &state,
+                task_config,
+                storage,
+                &conf.tenants_path(),
+                task_mgr::shutdown_token(),
+            )
+            .await;
+            info!("disk usage based eviction task finishing");
+            Ok(())
+        },
+    );
+
+    Ok(())
+}
+
+#[instrument(skip_all)]
+async fn disk_usage_eviction_task(
+    state: &State,
+    task_config: &DiskUsageEvictionTaskConfig,
+    storage: GenericRemoteStorage,
+    tenants_dir: &Path,
+    cancel: CancellationToken,
+) {
+    use crate::tenant::tasks::random_init_delay;
+    {
+        if random_init_delay(task_config.period, &cancel)
+            .await
+            .is_err()
+        {
+            info!("shutting down");
+            return;
+        }
+    }
+
+    let mut iteration_no = 0;
+    loop {
+        iteration_no += 1;
+        let start = Instant::now();
+
+        async {
+            let res = disk_usage_eviction_task_iteration(
+                state,
+                task_config,
+                &storage,
+                tenants_dir,
+                &cancel,
+            )
+            .await;
+
+            match res {
+                Ok(()) => {}
+                Err(e) => {
+                    // these stat failures are expected to be very rare
+                    warn!("iteration failed, unexpected error: {e:#}");
+                }
+            }
+        }
+        .instrument(tracing::info_span!("iteration", iteration_no))
+        .await;
+
+        let sleep_until = start + task_config.period;
+        tokio::select! {
+            _ = tokio::time::sleep_until(sleep_until) => {},
+            _ = cancel.cancelled() => {
+                info!("shutting down");
+                break
+            }
+        }
+    }
+}
+
+pub trait Usage: Clone + Copy + std::fmt::Debug {
+    fn has_pressure(&self) -> bool;
+    fn add_available_bytes(&mut self, bytes: u64);
+}
+
+async fn disk_usage_eviction_task_iteration(
+    state: &State,
+    task_config: &DiskUsageEvictionTaskConfig,
+    storage: &GenericRemoteStorage,
+    tenants_dir: &Path,
+    cancel: &CancellationToken,
+) -> anyhow::Result<()> {
+    let usage_pre = filesystem_level_usage::get(tenants_dir, task_config)
+        .context("get filesystem-level disk usage before evictions")?;
+    let res = disk_usage_eviction_task_iteration_impl(state, storage, usage_pre, cancel).await;
+    match res {
+        Ok(outcome) => {
+            debug!(?outcome, "disk_usage_eviction_iteration finished");
+            match outcome {
+                IterationOutcome::NoPressure | IterationOutcome::Cancelled => {
+                    // nothing to do, select statement below will handle things
+                }
+                IterationOutcome::Finished(outcome) => {
+                    // Verify with statvfs whether we made any real progress
+                    let after = filesystem_level_usage::get(tenants_dir, task_config)
+                        // It's quite unlikely to hit the error here. Keep the code simple and bail out.
+                        .context("get filesystem-level disk usage after evictions")?;
+
+                    debug!(?after, "disk usage");
+
+                    if after.has_pressure() {
+                        // Don't bother doing an out-of-order iteration here now.
+                        // In practice, the task period is set to a value in the tens-of-seconds range,
+                        // which will cause another iteration to happen soon enough.
+                        // TODO: deltas between the three different usages would be helpful,
+                        // consider MiB, GiB, TiB
+                        warn!(?outcome, ?after, "disk usage still high");
+                    } else {
+                        info!(?outcome, ?after, "disk usage pressure relieved");
+                    }
+                }
+            }
+        }
+        Err(e) => {
+            error!("disk_usage_eviction_iteration failed: {:#}", e);
+        }
+    }
+
+    Ok(())
+}
+
+#[derive(Debug, Serialize)]
+#[allow(clippy::large_enum_variant)]
+pub enum IterationOutcome<U> {
+    NoPressure,
+    Cancelled,
+    Finished(IterationOutcomeFinished<U>),
+}
+
+#[derive(Debug, Serialize)]
+pub struct IterationOutcomeFinished<U> {
+    /// The actual usage observed before we started the iteration.
+    before: U,
+    /// The expected value for `after`, according to internal accounting, after phase 1.
+    planned: PlannedUsage<U>,
+    /// The outcome of phase 2, where we actually do the evictions.
+    ///
+    /// If all layers that phase 1 planned to evict _can_ actually get evicted, this will
+    /// be the same as `planned`.
+    assumed: AssumedUsage<U>,
+}
+
+#[derive(Debug, Serialize)]
+struct AssumedUsage<U> {
+    /// The expected value for `after`, after phase 2.
+    projected_after: U,
+    /// The layers we failed to evict during phase 2.
+    failed: LayerCount,
+}
+
+#[derive(Debug, Serialize)]
+struct PlannedUsage<U> {
+    respecting_tenant_min_resident_size: U,
+    fallback_to_global_lru: Option<U>,
+}
+
+#[derive(Debug, Default, Serialize)]
+struct LayerCount {
+    file_sizes: u64,
+    count: usize,
+}
+
+pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
+    state: &State,
+    storage: &GenericRemoteStorage,
+    usage_pre: U,
+    cancel: &CancellationToken,
+) -> anyhow::Result<IterationOutcome<U>> {
+    // use tokio's mutex to get a Sync guard (instead of std::sync::Mutex)
+    let _g = state
+        .mutex
+        .try_lock()
+        .map_err(|_| anyhow::anyhow!("iteration is already executing"))?;
+
+    debug!(?usage_pre, "disk usage");
+
+    if !usage_pre.has_pressure() {
+        return Ok(IterationOutcome::NoPressure);
+    }
+
+    warn!(
+        ?usage_pre,
+        "running disk usage based eviction due to pressure"
+    );
+
+    let candidates = match collect_eviction_candidates(cancel).await? {
+        EvictionCandidates::Cancelled => {
+            return Ok(IterationOutcome::Cancelled);
+        }
+        EvictionCandidates::Finished(partitioned) => partitioned,
+    };
+
+    // Debug-log the list of candidates
+    let now = SystemTime::now();
+    for (i, (partition, candidate)) in candidates.iter().enumerate() {
+        debug!(
+            "cand {}/{}: size={}, no_access_for={}us, parition={:?}, tenant={} timeline={} layer={}",
+            i + 1,
+            candidates.len(),
+            candidate.layer.file_size(),
+            now.duration_since(candidate.last_activity_ts)
+                .unwrap()
+                .as_micros(),
+            partition,
+            candidate.layer.get_tenant_id(),
+            candidate.layer.get_timeline_id(),
+            candidate.layer.filename().file_name(),
+        );
+    }
+
+    // phase1: select victims to relieve pressure
+    //
+    // Walk through the list of candidates, until we have accumulated enough layers to get
+    // us back under the pressure threshold. 'usage_planned' is updated so that it tracks
+    // how much disk space would be used after evicting all the layers up to the current
+    // point in the list. The layers are collected in 'batched', grouped per timeline.
+    //
+    // If we get far enough in the list that we start to evict layers that are below
+    // the tenant's min-resident-size threshold, print a warning, and memorize the disk
+    // usage at that point, in 'usage_planned_min_resident_size_respecting'.
+    let mut batched: HashMap<_, Vec<Arc<dyn PersistentLayer>>> = HashMap::new();
+    let mut warned = None;
+    let mut usage_planned = usage_pre;
+    for (i, (partition, candidate)) in candidates.into_iter().enumerate() {
+        if !usage_planned.has_pressure() {
+            debug!(
+                no_candidates_evicted = i,
+                "took enough candidates for pressure to be relieved"
+            );
+            break;
+        }
+
+        if partition == MinResidentSizePartition::Below && warned.is_none() {
+            warn!(?usage_pre, ?usage_planned, candidate_no=i, "tenant_min_resident_size-respecting LRU would not relieve pressure, evicting more following global LRU policy");
+            warned = Some(usage_planned);
+        }
+
+        usage_planned.add_available_bytes(candidate.layer.file_size());
+
+        batched
+            .entry(TimelineKey(candidate.timeline))
+            .or_default()
+            .push(candidate.layer);
+    }
+
+    let usage_planned = match warned {
+        Some(respecting_tenant_min_resident_size) => PlannedUsage {
+            respecting_tenant_min_resident_size,
+            fallback_to_global_lru: Some(usage_planned),
+        },
+        None => PlannedUsage {
+            respecting_tenant_min_resident_size: usage_planned,
+            fallback_to_global_lru: None,
+        },
+    };
+    debug!(?usage_planned, "usage planned");
+
+    // phase2: evict victims batched by timeline
+
+    // After the loop, `usage_assumed` is the post-eviction usage,
+    // according to internal accounting.
+    let mut usage_assumed = usage_pre;
+    let mut evictions_failed = LayerCount::default();
+    for (timeline, batch) in batched {
+        let tenant_id = timeline.tenant_id;
+        let timeline_id = timeline.timeline_id;
+        let batch_size = batch.len();
+
+        debug!(%timeline_id, "evicting batch for timeline");
+
+        async {
+            let results = timeline.evict_layers(storage, &batch, cancel.clone()).await;
+
+            match results {
+                Err(e) => {
+                    warn!("failed to evict batch: {:#}", e);
+                }
+                Ok(results) => {
+                    assert_eq!(results.len(), batch.len());
+                    for (result, layer) in results.into_iter().zip(batch.iter()) {
+                        match result {
+                            Some(Ok(true)) => {
+                                usage_assumed.add_available_bytes(layer.file_size());
+                            }
+                            Some(Ok(false)) => {
+                                // this is:
+                                // - Replacement::{NotFound, Unexpected}
+                                // - it cannot be is_remote_layer, filtered already
+                                evictions_failed.file_sizes += layer.file_size();
+                                evictions_failed.count += 1;
+                            }
+                            None => {
+                                assert!(cancel.is_cancelled());
+                                return;
+                            }
+                            Some(Err(e)) => {
+                                // we really shouldn't be getting this, precondition failure
+                                error!("failed to evict layer: {:#}", e);
+                            }
+                        }
+                    }
+                }
+            }
+        }
+        .instrument(tracing::info_span!("evict_batch", %tenant_id, %timeline_id, batch_size))
+        .await;
+
+        if cancel.is_cancelled() {
+            return Ok(IterationOutcome::Cancelled);
+        }
+    }
+
+    Ok(IterationOutcome::Finished(IterationOutcomeFinished {
+        before: usage_pre,
+        planned: usage_planned,
+        assumed: AssumedUsage {
+            projected_after: usage_assumed,
+            failed: evictions_failed,
+        },
+    }))
+}
+
+#[derive(Clone)]
+struct EvictionCandidate {
+    timeline: Arc<Timeline>,
+    layer: Arc<dyn PersistentLayer>,
+    last_activity_ts: SystemTime,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
+enum MinResidentSizePartition {
+    Above,
+    Below,
+}
+
+enum EvictionCandidates {
+    Cancelled,
+    Finished(Vec<(MinResidentSizePartition, EvictionCandidate)>),
+}
+
+/// Gather the eviction candidates.
+///
+/// The returned `Ok(EvictionCandidates::Finished(candidates))` is sorted in eviction
+/// order. A caller that evicts in that order, until pressure is relieved, implements
+/// the eviction policy outlined in the module comment.
+///
+/// # Example
+///
+/// Imagine that there are two tenants, A and B, with five layers each, a-e.
+/// Each layer has size 100, and both tenant's min_resident_size is 150.
+/// The eviction order would be
+///
+/// ```text
+/// partition last_activity_ts    tenant/layer
+/// Above     18:30               A/c
+/// Above     19:00               A/b
+/// Above     18:29               B/c
+/// Above     19:05               B/b
+/// Above     20:00               B/a
+/// Above     20:03               A/a
+/// Below     20:30               A/d
+/// Below     20:40               B/d
+/// Below     20:45               B/e
+/// Below     20:58               A/e
+/// ```
+///
+/// Now, if we need to evict 300 bytes to relieve pressure, we'd evict `A/c, A/b, B/c`.
+/// They are all in the `Above` partition, so, we respected each tenant's min_resident_size.
+///
+/// But, if we need to evict 900 bytes to relieve pressure, we'd evict
+/// `A/c, A/b, B/c, B/b, B/a, A/a, A/d, B/d, B/e`, reaching into the `Below` partition
+/// after exhauting the `Above` partition.
+/// So, we did not respect each tenant's min_resident_size.
+async fn collect_eviction_candidates(
+    cancel: &CancellationToken,
+) -> anyhow::Result<EvictionCandidates> {
+    // get a snapshot of the list of tenants
+    let tenants = tenant::mgr::list_tenants()
+        .await
+        .context("get list of tenants")?;
+
+    let mut candidates = Vec::new();
+
+    for (tenant_id, _state) in &tenants {
+        if cancel.is_cancelled() {
+            return Ok(EvictionCandidates::Cancelled);
+        }
+        let tenant = match tenant::mgr::get_tenant(*tenant_id, true).await {
+            Ok(tenant) => tenant,
+            Err(e) => {
+                // this can happen if tenant has lifecycle transition after we fetched it
+                debug!("failed to get tenant: {e:#}");
+                continue;
+            }
+        };
+
+        // collect layers from all timelines in this tenant
+        //
+        // If one of the timelines becomes `!is_active()` during the iteration,
+        // for example because we're shutting down, then `max_layer_size` can be too small.
+        // That's OK. This code only runs under a disk pressure situation, and being
+        // a little unfair to tenants during shutdown in such a situation is tolerable.
+        let mut tenant_candidates = Vec::new();
+        let mut max_layer_size = 0;
+        for tl in tenant.list_timelines() {
+            if !tl.is_active() {
+                continue;
+            }
+            let info = tl.get_local_layers_for_disk_usage_eviction();
+            debug!(tenant_id=%tl.tenant_id, timeline_id=%tl.timeline_id, "timeline resident layers count: {}", info.resident_layers.len());
+            tenant_candidates.extend(
+                info.resident_layers
+                    .into_iter()
+                    .map(|layer_infos| (tl.clone(), layer_infos)),
+            );
+            max_layer_size = max_layer_size.max(info.max_layer_size.unwrap_or(0));
+
+            if cancel.is_cancelled() {
+                return Ok(EvictionCandidates::Cancelled);
+            }
+        }
+
+        // `min_resident_size` defaults to maximum layer file size of the tenant.
+        // This ensures that each tenant can have at least one layer resident at a given time,
+        // ensuring forward progress for a single Timeline::get in that tenant.
+        // It's a questionable heuristic since, usually, there are many Timeline::get
+        // requests going on for a tenant, and, at least in Neon prod, the median
+        // layer file size is much smaller than the compaction target size.
+        // We could be better here, e.g., sum of all L0 layers + most recent L1 layer.
+        // That's what's typically used by the various background loops.
+        //
+        // The default can be overriden with a fixed value in the tenant conf.
+        // A default override can be put in the default tenant conf in the pageserver.toml.
+        let min_resident_size = if let Some(s) = tenant.get_min_resident_size_override() {
+            debug!(
+                tenant_id=%tenant.tenant_id(),
+                overriden_size=s,
+                "using overridden min resident size for tenant"
+            );
+            s
+        } else {
+            debug!(
+                tenant_id=%tenant.tenant_id(),
+                max_layer_size,
+                "using max layer size as min_resident_size for tenant",
+            );
+            max_layer_size
+        };
+
+        // Sort layers most-recently-used first, then partition by
+        // cumsum above/below min_resident_size.
+        tenant_candidates
+            .sort_unstable_by_key(|(_, layer_info)| std::cmp::Reverse(layer_info.last_activity_ts));
+        let mut cumsum: i128 = 0;
+        for (timeline, layer_info) in tenant_candidates.into_iter() {
+            let file_size = layer_info.file_size();
+            let candidate = EvictionCandidate {
+                timeline,
+                last_activity_ts: layer_info.last_activity_ts,
+                layer: layer_info.layer,
+            };
+            let partition = if cumsum > min_resident_size as i128 {
+                MinResidentSizePartition::Above
+            } else {
+                MinResidentSizePartition::Below
+            };
+            candidates.push((partition, candidate));
+            cumsum += i128::from(file_size);
+        }
+    }
+
+    debug_assert!(MinResidentSizePartition::Above < MinResidentSizePartition::Below,
+        "as explained in the function's doc comment, layers that aren't in the tenant's min_resident_size are evicted first");
+    candidates
+        .sort_unstable_by_key(|(partition, candidate)| (*partition, candidate.last_activity_ts));
+
+    Ok(EvictionCandidates::Finished(candidates))
+}
+
+struct TimelineKey(Arc<Timeline>);
+
+impl PartialEq for TimelineKey {
+    fn eq(&self, other: &Self) -> bool {
+        Arc::ptr_eq(&self.0, &other.0)
+    }
+}
+
+impl Eq for TimelineKey {}
+
+impl std::hash::Hash for TimelineKey {
+    fn hash<H: std::hash::Hasher>(&self, state: &mut H) {
+        Arc::as_ptr(&self.0).hash(state);
+    }
+}
+
+impl std::ops::Deref for TimelineKey {
+    type Target = Timeline;
+
+    fn deref(&self) -> &Self::Target {
+        self.0.as_ref()
+    }
+}
+
+mod filesystem_level_usage {
+    use std::path::Path;
+
+    use anyhow::Context;
+
+    use crate::statvfs::Statvfs;
+
+    use super::DiskUsageEvictionTaskConfig;
+
+    #[derive(Debug, Clone, Copy)]
+    pub struct Usage<'a> {
+        config: &'a DiskUsageEvictionTaskConfig,
+
+        /// Filesystem capacity
+        total_bytes: u64,
+        /// Free filesystem space
+        avail_bytes: u64,
+    }
+
+    impl super::Usage for Usage<'_> {
+        fn has_pressure(&self) -> bool {
+            let usage_pct =
+                (100.0 * (1.0 - ((self.avail_bytes as f64) / (self.total_bytes as f64)))) as u64;
+
+            let pressures = [
+                (
+                    "min_avail_bytes",
+                    self.avail_bytes < self.config.min_avail_bytes,
+                ),
+                (
+                    "max_usage_pct",
+                    usage_pct > self.config.max_usage_pct.get() as u64,
+                ),
+            ];
+
+            pressures.into_iter().any(|(_, has_pressure)| has_pressure)
+        }
+
+        fn add_available_bytes(&mut self, bytes: u64) {
+            self.avail_bytes += bytes;
+        }
+    }
+
+    pub fn get<'a>(
+        tenants_dir: &Path,
+        config: &'a DiskUsageEvictionTaskConfig,
+    ) -> anyhow::Result<Usage<'a>> {
+        let mock_config = {
+            #[cfg(feature = "testing")]
+            {
+                config.mock_statvfs.as_ref()
+            }
+            #[cfg(not(feature = "testing"))]
+            {
+                None
+            }
+        };
+
+        let stat = Statvfs::get(tenants_dir, mock_config)
+            .context("statvfs failed, presumably directory got unlinked")?;
+
+        // https://unix.stackexchange.com/a/703650
+        let blocksize = if stat.fragment_size() > 0 {
+            stat.fragment_size()
+        } else {
+            stat.block_size()
+        };
+
+        // use blocks_available (b_avail) since, pageserver runs as unprivileged user
+        let avail_bytes = stat.blocks_available() * blocksize;
+        let total_bytes = stat.blocks() * blocksize;
+
+        Ok(Usage {
+            config,
+            total_bytes,
+            avail_bytes,
+        })
+    }
+}

pageserver/src/http/openapi_spec.yml

@@ -27,6 +27,31 @@ paths:
                   id:
                     type: integer
 
+  /v1/disk_usage_eviction/run:
+    put:
+      description: Do an iteration of disk-usage-based eviction to evict a given amount of disk space.
+      security: []
+      requestBody:
+        content:
+          application/json:
+            schema:
+              type: object
+              required:
+                - evict_bytes
+              properties:
+                evict_bytes:
+                  type: integer
+      responses:
+        "200":
+          description: |
+            The run completed.
+            This does not necessarily mean that we actually evicted `evict_bytes`.
+            Examine the returned object for detail, or, just watch the actual effect of the call using `du` or `df`.
+          content:
+            application/json:
+              schema:
+                type: object
+
   /v1/tenant/{tenant_id}:
     parameters:
       - name: tenant_id
@@ -189,13 +214,6 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/NotFoundError"
-        "412":
-          description: Tenant is missing
-          content:
-            application/json:
-              schema:
-                $ref: "#/components/schemas/PreconditionFailedError"
-
         "500":
           description: Generic operation error
           content:
@@ -873,9 +891,13 @@ components:
       type: object
       properties:
         tenant_specific_overrides:
-          $ref: "#/components/schemas/TenantConfigInfo"
+          type: object
+          schema:
+            $ref: "#/components/schemas/TenantConfigInfo"
         effective_config:
-          $ref: "#/components/schemas/TenantConfigInfo"
+          type: object
+          schema:
+            $ref: "#/components/schemas/TenantConfigInfo"
     TimelineInfo:
       type: object
       required:
@@ -961,13 +983,6 @@ components:
       properties:
         msg:
           type: string
-    PreconditionFailedError:
-      type: object
-      required:
-        - msg
-      properties:
-        msg:
-          type: string
 
 security:
   - JWT: []

pageserver/src/http/routes.rs

@@ -18,6 +18,7 @@ use super::models::{
     TimelineCreateRequest, TimelineGcRequest, TimelineInfo,
 };
 use crate::context::{DownloadBehavior, RequestContext};
+use crate::disk_usage_eviction_task;
 use crate::pgdatadir_mapping::LsnForTimestamp;
 use crate::task_mgr::TaskKind;
 use crate::tenant::config::TenantConfOpt;
@@ -48,6 +49,7 @@ struct State {
     auth: Option<Arc<JwtAuth>>,
     allowlist_routes: Vec<Uri>,
     remote_storage: Option<GenericRemoteStorage>,
+    disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
 }
 
 impl State {
@@ -55,6 +57,7 @@ impl State {
         conf: &'static PageServerConf,
         auth: Option<Arc<JwtAuth>>,
         remote_storage: Option<GenericRemoteStorage>,
+        disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
     ) -> anyhow::Result<Self> {
         let allowlist_routes = ["/v1/status", "/v1/doc", "/swagger.yml"]
             .iter()
@@ -65,6 +68,7 @@ impl State {
             auth,
             allowlist_routes,
             remote_storage,
+            disk_usage_eviction_state,
         })
     }
 }
@@ -148,11 +152,6 @@ impl From<crate::tenant::mgr::DeleteTimelineError> for ApiError {
     fn from(value: crate::tenant::mgr::DeleteTimelineError) -> Self {
         use crate::tenant::mgr::DeleteTimelineError::*;
         match value {
-            // Report Precondition failed so client can distinguish between
-            // "tenant is missing" case from "timeline is missing"
-            Tenant(TenantStateError::NotFound(..)) => {
-                ApiError::PreconditionFailed("Requested tenant is missing")
-            }
             Tenant(t) => ApiError::from(t),
             Timeline(t) => ApiError::from(t),
         }
@@ -775,6 +774,8 @@ async fn tenant_create_handler(mut request: Request<Body>) -> Result<Response<Bo
         );
     }
 
+    tenant_conf.min_resident_size_override = request_data.min_resident_size_override;
+
     let target_tenant_id = request_data
         .new_tenant_id
         .map(TenantId::from)
@@ -906,6 +907,8 @@ async fn update_tenant_config_handler(
         );
     }
 
+    tenant_conf.min_resident_size_override = request_data.min_resident_size_override;
+
     let state = get_state(&request);
     mgr::set_new_tenant_config(state.conf, tenant_conf, tenant_id)
         .instrument(info_span!("tenant_config", tenant = ?tenant_id))
@@ -914,6 +917,20 @@ async fn update_tenant_config_handler(
     json_response(StatusCode::OK, ())
 }
 
+/// Testing helper to transition a tenant to [`crate::tenant::TenantState::Broken`].
+#[cfg(feature = "testing")]
+async fn handle_tenant_break(r: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&r, "tenant_id")?;
+
+    let tenant = crate::tenant::mgr::get_tenant(tenant_id, true)
+        .await
+        .map_err(|_| ApiError::Conflict(String::from("no active tenant found")))?;
+
+    tenant.set_broken("broken from test");
+
+    json_response(StatusCode::OK, ())
+}
+
 #[cfg(feature = "testing")]
 async fn failpoints_handler(mut request: Request<Body>) -> Result<Response<Body>, ApiError> {
     if !fail::has_failpoints() {
@@ -1063,6 +1080,89 @@ async fn always_panic_handler(req: Request<Body>) -> Result<Response<Body>, ApiE
     json_response(StatusCode::NO_CONTENT, ())
 }
 
+async fn disk_usage_eviction_run(mut r: Request<Body>) -> Result<Response<Body>, ApiError> {
+    check_permission(&r, None)?;
+
+    #[derive(Debug, Clone, Copy, serde::Serialize, serde::Deserialize)]
+    struct Config {
+        /// How many bytes to evict before reporting that pressure is relieved.
+        evict_bytes: u64,
+    }
+
+    #[derive(Debug, Clone, Copy, serde::Serialize)]
+    struct Usage {
+        // remains unchanged after instantiation of the struct
+        config: Config,
+        // updated by `add_available_bytes`
+        freed_bytes: u64,
+    }
+
+    impl crate::disk_usage_eviction_task::Usage for Usage {
+        fn has_pressure(&self) -> bool {
+            self.config.evict_bytes > self.freed_bytes
+        }
+
+        fn add_available_bytes(&mut self, bytes: u64) {
+            self.freed_bytes += bytes;
+        }
+    }
+
+    let config = json_request::<Config>(&mut r)
+        .await
+        .map_err(|_| ApiError::BadRequest(anyhow::anyhow!("invalid JSON body")))?;
+
+    let usage = Usage {
+        config,
+        freed_bytes: 0,
+    };
+
+    use crate::task_mgr::MGMT_REQUEST_RUNTIME;
+
+    let (tx, rx) = tokio::sync::oneshot::channel();
+
+    let state = get_state(&r);
+
+    let Some(storage) = state.remote_storage.clone() else {
+        return Err(ApiError::InternalServerError(anyhow::anyhow!(
+            "remote storage not configured, cannot run eviction iteration"
+        )))
+    };
+
+    let state = state.disk_usage_eviction_state.clone();
+
+    let cancel = CancellationToken::new();
+    let child_cancel = cancel.clone();
+    let _g = cancel.drop_guard();
+
+    crate::task_mgr::spawn(
+        MGMT_REQUEST_RUNTIME.handle(),
+        TaskKind::DiskUsageEviction,
+        None,
+        None,
+        "ondemand disk usage eviction",
+        false,
+        async move {
+            let res = crate::disk_usage_eviction_task::disk_usage_eviction_task_iteration_impl(
+                &state,
+                &storage,
+                usage,
+                &child_cancel,
+            )
+            .await;
+
+            info!(?res, "disk_usage_eviction_task_iteration_impl finished");
+
+            let _ = tx.send(res);
+            Ok(())
+        }
+        .in_current_span(),
+    );
+
+    let response = rx.await.unwrap().map_err(ApiError::InternalServerError)?;
+
+    json_response(StatusCode::OK, response)
+}
+
 async fn handler_404(_: Request<Body>) -> Result<Response<Body>, ApiError> {
     json_response(
         StatusCode::NOT_FOUND,
@@ -1075,6 +1175,7 @@ pub fn make_router(
     launch_ts: &'static LaunchTimestamp,
     auth: Option<Arc<JwtAuth>>,
     remote_storage: Option<GenericRemoteStorage>,
+    disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
 ) -> anyhow::Result<RouterBuilder<hyper::Body, ApiError>> {
     let spec = include_bytes!("openapi_spec.yml");
     let mut router = attach_openapi_ui(endpoint::make_router(), spec, "/swagger.yml", "/v1/doc");
@@ -1119,7 +1220,8 @@ pub fn make_router(
 
     Ok(router
         .data(Arc::new(
-            State::new(conf, auth, remote_storage).context("Failed to initialize router state")?,
+            State::new(conf, auth, remote_storage, disk_usage_eviction_state)
+                .context("Failed to initialize router state")?,
         ))
         .get("/v1/status", |r| RequestSpan(status_handler).handle(r))
         .put(
@@ -1200,6 +1302,13 @@ pub fn make_router(
             "/v1/tenant/:tenant_id/timeline/:timeline_id/layer/:layer_file_name",
             |r| RequestSpan(evict_timeline_layer_handler).handle(r),
         )
+        .put("/v1/disk_usage_eviction/run", |r| {
+            RequestSpan(disk_usage_eviction_run).handle(r)
+        })
+        .put(
+            "/v1/tenant/:tenant_id/break",
+            testing_api!("set tenant state to broken", handle_tenant_break),
+        )
         .get("/v1/panic", |r| RequestSpan(always_panic_handler).handle(r))
         .any(handler_404))
 }

pageserver/src/lib.rs

@@ -4,6 +4,7 @@ pub mod broker_client;
 pub mod config;
 pub mod consumption_metrics;
 pub mod context;
+pub mod disk_usage_eviction_task;
 pub mod http;
 pub mod import_datadir;
 pub mod keyspace;
@@ -12,6 +13,7 @@ pub mod page_cache;
 pub mod page_service;
 pub mod pgdatadir_mapping;
 pub mod repository;
+pub(crate) mod statvfs;
 pub mod task_mgr;
 pub mod tenant;
 pub mod trace;

pageserver/src/page_service.rs

@@ -27,7 +27,6 @@ use pq_proto::FeStartupPacket;
 use pq_proto::{BeMessage, FeMessage, RowDescriptor};
 use std::io;
 use std::net::TcpListener;
-use std::pin::pin;
 use std::str;
 use std::str::FromStr;
 use std::sync::Arc;
@@ -467,7 +466,8 @@ impl PageServerHandler {
         pgb.write_message_noflush(&BeMessage::CopyInResponse)?;
         pgb.flush().await?;
 
-        let mut copyin_reader = pin!(StreamReader::new(copyin_stream(pgb)));
+        let copyin_reader = StreamReader::new(copyin_stream(pgb));
+        tokio::pin!(copyin_reader);
         timeline
             .import_basebackup_from_tar(&mut copyin_reader, base_lsn, &ctx)
             .await?;
@@ -512,7 +512,8 @@ impl PageServerHandler {
         info!("importing wal");
         pgb.write_message_noflush(&BeMessage::CopyInResponse)?;
         pgb.flush().await?;
-        let mut copyin_reader = pin!(StreamReader::new(copyin_stream(pgb)));
+        let copyin_reader = StreamReader::new(copyin_stream(pgb));
+        tokio::pin!(copyin_reader);
         import_wal_from_tar(&timeline, &mut copyin_reader, start_lsn, end_lsn, &ctx).await?;
         info!("wal import complete");
 

pageserver/src/statvfs.rs

@@ -0,0 +1,150 @@
+//! Wrapper around nix::sys::statvfs::Statvfs that allows for mocking.
+
+use std::path::Path;
+
+pub enum Statvfs {
+    Real(nix::sys::statvfs::Statvfs),
+    Mock(mock::Statvfs),
+}
+
+// NB: on macOS, the block count type of struct statvfs is u32.
+// The workaround seems to be to use the non-standard statfs64 call.
+// Sincce it should only be a problem on > 2TiB disks, let's ignore
+// the problem for now and upcast to u64.
+impl Statvfs {
+    pub fn get(tenants_dir: &Path, mocked: Option<&mock::Behavior>) -> nix::Result<Self> {
+        if let Some(mocked) = mocked {
+            Ok(Statvfs::Mock(mock::get(tenants_dir, mocked)?))
+        } else {
+            Ok(Statvfs::Real(nix::sys::statvfs::statvfs(tenants_dir)?))
+        }
+    }
+
+    // NB: allow() because the block count type is u32 on macOS.
+    #[allow(clippy::useless_conversion)]
+    pub fn blocks(&self) -> u64 {
+        match self {
+            Statvfs::Real(stat) => u64::try_from(stat.blocks()).unwrap(),
+            Statvfs::Mock(stat) => stat.blocks,
+        }
+    }
+
+    // NB: allow() because the block count type is u32 on macOS.
+    #[allow(clippy::useless_conversion)]
+    pub fn blocks_available(&self) -> u64 {
+        match self {
+            Statvfs::Real(stat) => u64::try_from(stat.blocks_available()).unwrap(),
+            Statvfs::Mock(stat) => stat.blocks_available,
+        }
+    }
+
+    pub fn fragment_size(&self) -> u64 {
+        match self {
+            Statvfs::Real(stat) => stat.fragment_size(),
+            Statvfs::Mock(stat) => stat.fragment_size,
+        }
+    }
+
+    pub fn block_size(&self) -> u64 {
+        match self {
+            Statvfs::Real(stat) => stat.block_size(),
+            Statvfs::Mock(stat) => stat.block_size,
+        }
+    }
+}
+
+pub mod mock {
+    use anyhow::Context;
+    use regex::Regex;
+    use std::path::Path;
+    use tracing::log::info;
+
+    #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+    #[serde(tag = "type")]
+    pub enum Behavior {
+        Success {
+            blocksize: u64,
+            total_blocks: u64,
+            name_filter: Option<utils::serde_regex::Regex>,
+        },
+        Failure {
+            mocked_error: MockedError,
+        },
+    }
+
+    #[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+    #[allow(clippy::upper_case_acronyms)]
+    pub enum MockedError {
+        EIO,
+    }
+
+    impl From<MockedError> for nix::Error {
+        fn from(e: MockedError) -> Self {
+            match e {
+                MockedError::EIO => nix::Error::EIO,
+            }
+        }
+    }
+
+    pub fn get(tenants_dir: &Path, behavior: &Behavior) -> nix::Result<Statvfs> {
+        info!("running mocked statvfs");
+
+        match behavior {
+            Behavior::Success {
+                blocksize,
+                total_blocks,
+                ref name_filter,
+            } => {
+                let used_bytes = walk_dir_disk_usage(tenants_dir, name_filter.as_deref()).unwrap();
+
+                // round it up to the nearest block multiple
+                let used_blocks = (used_bytes + (blocksize - 1)) / blocksize;
+
+                if used_blocks > *total_blocks {
+                    panic!(
+                        "mocking error: used_blocks > total_blocks: {used_blocks} > {total_blocks}"
+                    );
+                }
+
+                let avail_blocks = total_blocks - used_blocks;
+
+                Ok(Statvfs {
+                    blocks: *total_blocks,
+                    blocks_available: avail_blocks,
+                    fragment_size: *blocksize,
+                    block_size: *blocksize,
+                })
+            }
+            Behavior::Failure { mocked_error } => Err((*mocked_error).into()),
+        }
+    }
+
+    fn walk_dir_disk_usage(path: &Path, name_filter: Option<&Regex>) -> anyhow::Result<u64> {
+        let mut total = 0;
+        for entry in walkdir::WalkDir::new(path) {
+            let entry = entry?;
+            if !entry.file_type().is_file() {
+                continue;
+            }
+            if !name_filter
+                .as_ref()
+                .map(|filter| filter.is_match(entry.file_name().to_str().unwrap()))
+                .unwrap_or(true)
+            {
+                continue;
+            }
+            total += entry
+                .metadata()
+                .with_context(|| format!("get metadata of {:?}", entry.path()))?
+                .len();
+        }
+        Ok(total)
+    }
+
+    pub struct Statvfs {
+        pub blocks: u64,
+        pub blocks_available: u64,
+        pub fragment_size: u64,
+        pub block_size: u64,
+    }
+}

pageserver/src/task_mgr.rs

@@ -234,6 +234,9 @@ pub enum TaskKind {
     // Eviction. One per timeline.
     Eviction,
 
+    /// See [`crate::disk_usage_eviction_task`].
+    DiskUsageEviction,
+
     // Initial logical size calculation
     InitialLogicalSizeCalculation,
 

pageserver/src/tenant.rs

@@ -46,7 +46,6 @@ use std::time::{Duration, Instant};
 use self::config::TenantConf;
 use self::metadata::TimelineMetadata;
 use self::remote_timeline_client::RemoteTimelineClient;
-use self::timeline::EvictionTaskTenantState;
 use crate::config::PageServerConf;
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::import_datadir;
@@ -95,7 +94,7 @@ mod timeline;
 
 pub mod size;
 
-pub use timeline::{PageReconstructError, Timeline};
+pub use timeline::{LocalLayerInfoForDiskUsageEviction, PageReconstructError, Timeline};
 
 // re-export this function so that page_cache.rs can use it.
 pub use crate::tenant::ephemeral_file::writeback as writeback_ephemeral_file;
@@ -143,8 +142,6 @@ pub struct Tenant {
     /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
     cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
     cached_synthetic_tenant_size: Arc<AtomicU64>,
-
-    eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
 }
 
 /// A timeline with some of its files on disk, being initialized.
@@ -1706,6 +1703,13 @@ impl Tenant {
             .unwrap_or(self.conf.default_tenant_conf.trace_read_requests)
     }
 
+    pub fn get_min_resident_size_override(&self) -> Option<u64> {
+        let tenant_conf = self.tenant_conf.read().unwrap();
+        tenant_conf
+            .min_resident_size_override
+            .or(self.conf.default_tenant_conf.min_resident_size_override)
+    }
+
     pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
         *self.tenant_conf.write().unwrap() = new_tenant_conf;
     }
@@ -1784,7 +1788,6 @@ impl Tenant {
             state,
             cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
             cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
-            eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
         }
     }
 
@@ -2783,6 +2786,7 @@ pub mod harness {
                 max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
                 trace_read_requests: Some(tenant_conf.trace_read_requests),
                 eviction_policy: Some(tenant_conf.eviction_policy),
+                min_resident_size_override: tenant_conf.min_resident_size_override,
             }
         }
     }

pageserver/src/tenant/config.rs

@@ -92,6 +92,7 @@ pub struct TenantConf {
     pub max_lsn_wal_lag: NonZeroU64,
     pub trace_read_requests: bool,
     pub eviction_policy: EvictionPolicy,
+    pub min_resident_size_override: Option<u64>,
 }
 
 /// Same as TenantConf, but this struct preserves the information about
@@ -159,6 +160,10 @@ pub struct TenantConfOpt {
     #[serde(skip_serializing_if = "Option::is_none")]
     #[serde(default)]
     pub eviction_policy: Option<EvictionPolicy>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[serde(default)]
+    pub min_resident_size_override: Option<u64>,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
@@ -220,6 +225,9 @@ impl TenantConfOpt {
                 .trace_read_requests
                 .unwrap_or(global_conf.trace_read_requests),
             eviction_policy: self.eviction_policy.unwrap_or(global_conf.eviction_policy),
+            min_resident_size_override: self
+                .min_resident_size_override
+                .or(global_conf.min_resident_size_override),
         }
     }
 }
@@ -251,6 +259,7 @@ impl Default for TenantConf {
                 .expect("cannot parse default max walreceiver Lsn wal lag"),
             trace_read_requests: false,
             eviction_policy: EvictionPolicy::NoEviction,
+            min_resident_size_override: None,
         }
     }
 }

pageserver/src/tenant/size.rs

@@ -6,7 +6,6 @@ use std::sync::Arc;
 use anyhow::{bail, Context};
 use tokio::sync::oneshot::error::RecvError;
 use tokio::sync::Semaphore;
-use tokio_util::sync::CancellationToken;
 
 use crate::context::RequestContext;
 use crate::pgdatadir_mapping::CalculateLogicalSizeError;
@@ -353,10 +352,6 @@ async fn fill_logical_sizes(
     // our advantage with `?` error handling.
     let mut joinset = tokio::task::JoinSet::new();
 
-    let cancel = tokio_util::sync::CancellationToken::new();
-    // be sure to cancel all spawned tasks if we are dropped
-    let _dg = cancel.clone().drop_guard();
-
     // For each point that would benefit from having a logical size available,
     // spawn a Task to fetch it, unless we have it cached already.
     for seg in segments.iter() {
@@ -378,7 +373,6 @@ async fn fill_logical_sizes(
                     timeline,
                     lsn,
                     ctx,
-                    cancel.child_token(),
                 ));
             }
             e.insert(cached_size);
@@ -483,14 +477,13 @@ async fn calculate_logical_size(
     timeline: Arc<crate::tenant::Timeline>,
     lsn: utils::lsn::Lsn,
     ctx: RequestContext,
-    cancel: CancellationToken,
 ) -> Result<TimelineAtLsnSizeResult, RecvError> {
     let _permit = tokio::sync::Semaphore::acquire_owned(limit)
         .await
         .expect("global semaphore should not had been closed");
 
     let size_res = timeline
-        .spawn_ondemand_logical_size_calculation(lsn, ctx, cancel)
+        .spawn_ondemand_logical_size_calculation(lsn, ctx)
         .instrument(info_span!("spawn_ondemand_logical_size_calculation"))
         .await?;
     Ok(TimelineAtLsnSizeResult(timeline, lsn, size_res))

pageserver/src/tenant/storage_layer.rs

@@ -121,10 +121,10 @@ struct LayerAccessStatsInner {
 }
 
 #[derive(Debug, Clone, Copy)]
-pub(super) struct LayerAccessStatFullDetails {
-    pub(super) when: SystemTime,
-    pub(super) task_kind: TaskKind,
-    pub(super) access_kind: LayerAccessKind,
+pub(crate) struct LayerAccessStatFullDetails {
+    pub(crate) when: SystemTime,
+    pub(crate) task_kind: TaskKind,
+    pub(crate) access_kind: LayerAccessKind,
 }
 
 #[derive(Clone, Copy, strum_macros::EnumString)]
@@ -255,7 +255,7 @@ impl LayerAccessStats {
         ret
     }
 
-    pub(super) fn most_recent_access_or_residence_event(
+    fn most_recent_access_or_residence_event(
         &self,
     ) -> Either<LayerAccessStatFullDetails, LayerResidenceEvent> {
         let locked = self.0.lock().unwrap();
@@ -268,6 +268,13 @@ impl LayerAccessStats {
             }
         }
     }
+
+    pub(crate) fn latest_activity(&self) -> SystemTime {
+        match self.most_recent_access_or_residence_event() {
+            Either::Left(mra) => mra.when,
+            Either::Right(re) => re.timestamp,
+        }
+    }
 }
 
 /// Supertrait of the [`Layer`] trait that captures the bare minimum interface

pageserver/src/tenant/tasks.rs

@@ -244,14 +244,12 @@ pub(crate) async fn random_init_delay(
 ) -> Result<(), Cancelled> {
     use rand::Rng;
 
+    if period == Duration::ZERO {
+        return Ok(());
+    }
+
     let d = {
         let mut rng = rand::thread_rng();
-
-        // gen_range asserts that the range cannot be empty, which it could be because period can
-        // be set to zero to disable gc or compaction, so lets set it to be at least 10s.
-        let period = std::cmp::max(period, Duration::from_secs(10));
-
-        // semi-ok default as the source of jitter
         rng.gen_range(Duration::ZERO..=period)
     };
 

pageserver/src/tenant/timeline.rs

@@ -13,6 +13,7 @@ use pageserver_api::models::{
     DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskSpawnRequest,
     DownloadRemoteLayersTaskState, LayerMapInfo, LayerResidenceStatus, TimelineState,
 };
+use remote_storage::GenericRemoteStorage;
 use tokio::sync::{oneshot, watch, Semaphore, TryAcquireError};
 use tokio_util::sync::CancellationToken;
 use tracing::*;
@@ -24,7 +25,6 @@ use std::collections::HashMap;
 use std::fs;
 use std::ops::{Deref, Range};
 use std::path::{Path, PathBuf};
-use std::pin::pin;
 use std::sync::atomic::{AtomicI64, Ordering as AtomicOrdering};
 use std::sync::{Arc, Mutex, MutexGuard, RwLock, Weak};
 use std::time::{Duration, Instant, SystemTime};
@@ -72,7 +72,6 @@ use crate::ZERO_PAGE;
 use crate::{is_temporary, task_mgr};
 use walreceiver::spawn_connection_manager_task;
 
-pub(super) use self::eviction_task::EvictionTaskTenantState;
 use self::eviction_task::EvictionTaskTimelineState;
 
 use super::layer_map::BatchedUpdates;
@@ -678,7 +677,8 @@ impl Timeline {
 
             let mut failed = 0;
 
-            let mut cancelled = pin!(task_mgr::shutdown_watcher());
+            let cancelled = task_mgr::shutdown_watcher();
+            tokio::pin!(cancelled);
 
             loop {
                 tokio::select! {
@@ -957,6 +957,25 @@ impl Timeline {
         }
     }
 
+    /// Evict a batch of layers.
+    ///
+    /// GenericRemoteStorage reference is required as a witness[^witness_article] for "remote storage is configured."
+    ///
+    /// [^witness_article]: https://willcrichton.net/rust-api-type-patterns/witnesses.html
+    pub async fn evict_layers(
+        &self,
+        _: &GenericRemoteStorage,
+        layers_to_evict: &[Arc<dyn PersistentLayer>],
+        cancel: CancellationToken,
+    ) -> anyhow::Result<Vec<Option<anyhow::Result<bool>>>> {
+        let remote_client = self.remote_client.clone().expect(
+            "GenericRemoteStorage is configured, so timeline must have RemoteTimelineClient",
+        );
+
+        self.evict_layer_batch(&remote_client, layers_to_evict, cancel)
+            .await
+    }
+
     /// Evict multiple layers at once, continuing through errors.
     ///
     /// Try to evict the given `layers_to_evict` by
@@ -994,6 +1013,15 @@ impl Timeline {
         // now lock out layer removal (compaction, gc, timeline deletion)
         let layer_removal_guard = self.layer_removal_cs.lock().await;
 
+        {
+            // to avoid racing with detach and delete_timeline
+            let state = self.current_state();
+            anyhow::ensure!(
+                state == TimelineState::Active,
+                "timeline is not active but {state:?}"
+            );
+        }
+
         // start the batch update
         let mut layer_map = self.layers.write().unwrap();
         let mut batch_updates = layer_map.batch_update();
@@ -1027,6 +1055,8 @@ impl Timeline {
         use super::layer_map::Replacement;
 
         if local_layer.is_remote_layer() {
+            // TODO(issue #3851): consider returning an err here instead of false,
+            // which is the same out the match later
             return Ok(false);
         }
 
@@ -1738,11 +1768,8 @@ impl Timeline {
             false,
             // NB: don't log errors here, task_mgr will do that.
             async move {
-                // no cancellation here, because nothing really waits for this to complete compared
-                // to spawn_ondemand_logical_size_calculation.
-                let cancel = CancellationToken::new();
                 let calculated_size = match self_clone
-                    .logical_size_calculation_task(lsn, &background_ctx, cancel)
+                    .logical_size_calculation_task(lsn, &background_ctx)
                     .await
                 {
                     Ok(s) => s,
@@ -1797,7 +1824,6 @@ impl Timeline {
         self: &Arc<Self>,
         lsn: Lsn,
         ctx: RequestContext,
-        cancel: CancellationToken,
     ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
         let (sender, receiver) = oneshot::channel();
         let self_clone = Arc::clone(self);
@@ -1817,9 +1843,7 @@ impl Timeline {
             "ondemand logical size calculation",
             false,
             async move {
-                let res = self_clone
-                    .logical_size_calculation_task(lsn, &ctx, cancel)
-                    .await;
+                let res = self_clone.logical_size_calculation_task(lsn, &ctx).await;
                 let _ = sender.send(res).ok();
                 Ok(()) // Receiver is responsible for handling errors
             },
@@ -1832,18 +1856,18 @@ impl Timeline {
         self: &Arc<Self>,
         lsn: Lsn,
         ctx: &RequestContext,
-        cancel: CancellationToken,
     ) -> Result<u64, CalculateLogicalSizeError> {
         let mut timeline_state_updates = self.subscribe_for_state_updates();
         let self_calculation = Arc::clone(self);
+        let cancel = CancellationToken::new();
 
-        let mut calculation = pin!(async {
+        let calculation = async {
             let cancel = cancel.child_token();
             let ctx = ctx.attached_child();
             self_calculation
                 .calculate_logical_size(lsn, cancel, &ctx)
                 .await
-        });
+        };
         let timeline_state_cancellation = async {
             loop {
                 match timeline_state_updates.changed().await {
@@ -1872,6 +1896,7 @@ impl Timeline {
             "aborted because task_mgr shutdown requested".to_string()
         };
 
+        tokio::pin!(calculation);
         loop {
             tokio::select! {
                 res = &mut calculation => { return res }
@@ -4012,6 +4037,67 @@ impl Timeline {
     }
 }
 
+pub struct DiskUsageEvictionInfo {
+    /// Timeline's largest layer (remote or resident)
+    pub max_layer_size: Option<u64>,
+    /// Timeline's resident layers
+    pub resident_layers: Vec<LocalLayerInfoForDiskUsageEviction>,
+}
+
+pub struct LocalLayerInfoForDiskUsageEviction {
+    pub layer: Arc<dyn PersistentLayer>,
+    pub last_activity_ts: SystemTime,
+}
+
+impl std::fmt::Debug for LocalLayerInfoForDiskUsageEviction {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        // format the tv_sec, tv_nsec into rfc3339 in case someone is looking at it
+        // having to allocate a string to this is bad, but it will rarely be formatted
+        let ts = chrono::DateTime::<chrono::Utc>::from(self.last_activity_ts);
+        let ts = ts.to_rfc3339_opts(chrono::SecondsFormat::Nanos, true);
+        f.debug_struct("LocalLayerInfoForDiskUsageEviction")
+            .field("layer", &self.layer)
+            .field("last_activity", &ts)
+            .finish()
+    }
+}
+
+impl LocalLayerInfoForDiskUsageEviction {
+    pub fn file_size(&self) -> u64 {
+        self.layer.file_size()
+    }
+}
+
+impl Timeline {
+    pub(crate) fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
+        let layers = self.layers.read().unwrap();
+
+        let mut max_layer_size: Option<u64> = None;
+        let mut resident_layers = Vec::new();
+
+        for l in layers.iter_historic_layers() {
+            let file_size = l.file_size();
+            max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
+
+            if l.is_remote_layer() {
+                continue;
+            }
+
+            let last_activity_ts = l.access_stats().latest_activity();
+
+            resident_layers.push(LocalLayerInfoForDiskUsageEviction {
+                layer: l,
+                last_activity_ts,
+            });
+        }
+
+        DiskUsageEvictionInfo {
+            max_layer_size,
+            resident_layers,
+        }
+    }
+}
+
 type TraversalPathItem = (
     ValueReconstructResult,
     Lsn,

pageserver/src/tenant/timeline/eviction_task.rs

@@ -14,13 +14,11 @@
 //!
 //! See write-up on restart on-demand download spike: <https://gist.github.com/problame/2265bf7b8dc398be834abfead36c76b5>
 use std::{
-    collections::HashMap,
     ops::ControlFlow,
     sync::Arc,
     time::{Duration, SystemTime},
 };
 
-use either::Either;
 use tokio::time::Instant;
 use tokio_util::sync::CancellationToken;
 use tracing::{debug, error, info, instrument, warn};
@@ -31,7 +29,6 @@ use crate::{
     tenant::{
         config::{EvictionPolicy, EvictionPolicyLayerAccessThreshold},
         storage_layer::PersistentLayer,
-        Tenant,
     },
 };
 
@@ -39,12 +36,7 @@ use super::Timeline;
 
 #[derive(Default)]
 pub struct EvictionTaskTimelineState {
-    last_layer_access_imitation: Option<tokio::time::Instant>,
-}
-
-#[derive(Default)]
-pub struct EvictionTaskTenantState {
-    last_layer_access_imitation: Option<Instant>,
+    last_refresh_required_in_restart: Option<tokio::time::Instant>,
 }
 
 impl Timeline {
@@ -134,35 +126,6 @@ impl Timeline {
     ) -> ControlFlow<()> {
         let now = SystemTime::now();
 
-        // If we evict layers but keep cached values derived from those layers, then
-        // we face a storm of on-demand downloads after pageserver restart.
-        // The reason is that the restart empties the caches, and so, the values
-        // need to be re-computed by accessing layers, which we evicted while the
-        // caches were filled.
-        //
-        // Solutions here would be one of the following:
-        // 1. Have a persistent cache.
-        // 2. Count every access to a cached value to the access stats of all layers
-        //    that were accessed to compute the value in the first place.
-        // 3. Invalidate the caches at a period of < p.threshold/2, so that the values
-        //    get re-computed from layers, thereby counting towards layer access stats.
-        // 4. Make the eviction task imitate the layer accesses that typically hit caches.
-        //
-        // We follow approach (4) here because in Neon prod deployment:
-        // - page cache is quite small => high churn => low hit rate
-        //   => eviction gets correct access stats
-        // - value-level caches such as logical size & repatition have a high hit rate,
-        //   especially for inactive tenants
-        //   => eviction sees zero accesses for these
-        //   => they cause the on-demand download storm on pageserver restart
-        //
-        // We should probably move to persistent caches in the future, or avoid
-        // having inactive tenants attached to pageserver in the first place.
-        match self.imitate_layer_accesses(p, cancel, ctx).await {
-            ControlFlow::Break(()) => return ControlFlow::Break(()),
-            ControlFlow::Continue(()) => (),
-        }
-
         #[allow(dead_code)]
         #[derive(Debug, Default)]
         struct EvictionStats {
@@ -173,6 +136,27 @@ impl Timeline {
             skipped_for_shutdown: usize,
         }
 
+        // what we want is to invalidate any caches which haven't been accessed for `p.threshold`,
+        // but we cannot actually do it for current limitations except by restarting pageserver. we
+        // just recompute the values which would be recomputed on startup.
+        //
+        // for active tenants this will likely materialized page cache or in-memory layers. for
+        // inactive tenants it will refresh the last_access timestamps so that we will not evict
+        // and re-download on restart these layers.
+        let mut state = self.eviction_task_timeline_state.lock().await;
+        match state.last_refresh_required_in_restart {
+            Some(ts) if ts.elapsed() < p.threshold => { /* no need to run */ }
+            _ => {
+                self.refresh_layers_required_in_restart(cancel, ctx).await;
+                state.last_refresh_required_in_restart = Some(tokio::time::Instant::now())
+            }
+        }
+        drop(state);
+
+        if cancel.is_cancelled() {
+            return ControlFlow::Break(());
+        }
+
         let mut stats = EvictionStats::default();
         // Gather layers for eviction.
         // NB: all the checks can be invalidated as soon as we release the layer map lock.
@@ -185,13 +169,7 @@ impl Timeline {
                 if hist_layer.is_remote_layer() {
                     continue;
                 }
-                let last_activity_ts = match hist_layer
-                    .access_stats()
-                    .most_recent_access_or_residence_event()
-                {
-                    Either::Left(mra) => mra.when,
-                    Either::Right(re) => re.timestamp,
-                };
+                let last_activity_ts = hist_layer.access_stats().latest_activity();
                 let no_activity_for = match now.duration_since(last_activity_ts) {
                     Ok(d) => d,
                     Err(_e) => {
@@ -276,55 +254,8 @@ impl Timeline {
         ControlFlow::Continue(())
     }
 
-    async fn imitate_layer_accesses(
-        &self,
-        p: &EvictionPolicyLayerAccessThreshold,
-        cancel: &CancellationToken,
-        ctx: &RequestContext,
-    ) -> ControlFlow<()> {
-        let mut state = self.eviction_task_timeline_state.lock().await;
-        match state.last_layer_access_imitation {
-            Some(ts) if ts.elapsed() < p.threshold => { /* no need to run */ }
-            _ => {
-                self.imitate_timeline_cached_layer_accesses(cancel, ctx)
-                    .await;
-                state.last_layer_access_imitation = Some(tokio::time::Instant::now())
-            }
-        }
-        drop(state);
-
-        if cancel.is_cancelled() {
-            return ControlFlow::Break(());
-        }
-
-        // This task is timeline-scoped, but the synthetic size calculation is tenant-scoped.
-        // Make one of the tenant's timelines draw the short straw and run the calculation.
-        // The others wait until the calculation is done so that they take into account the
-        // imitated accesses that the winner made.
-        let Ok(tenant) = crate::tenant::mgr::get_tenant(self.tenant_id, true).await else {
-            // likely, we're shutting down
-            return ControlFlow::Break(());
-        };
-        let mut state = tenant.eviction_task_tenant_state.lock().await;
-        match state.last_layer_access_imitation {
-            Some(ts) if ts.elapsed() < p.threshold => { /* no need to run */ }
-            _ => {
-                self.imitate_synthetic_size_calculation_worker(&tenant, ctx, cancel)
-                    .await;
-                state.last_layer_access_imitation = Some(tokio::time::Instant::now());
-            }
-        }
-        drop(state);
-
-        if cancel.is_cancelled() {
-            return ControlFlow::Break(());
-        }
-
-        ControlFlow::Continue(())
-    }
-
     /// Recompute the values which would cause on-demand downloads during restart.
-    async fn imitate_timeline_cached_layer_accesses(
+    async fn refresh_layers_required_in_restart(
         &self,
         cancel: &CancellationToken,
         ctx: &RequestContext,
@@ -358,61 +289,4 @@ impl Timeline {
             }
         }
     }
-
-    // Imitate the synthetic size calculation done by the consumption_metrics module.
-    async fn imitate_synthetic_size_calculation_worker(
-        &self,
-        tenant: &Arc<Tenant>,
-        ctx: &RequestContext,
-        cancel: &CancellationToken,
-    ) {
-        if self.conf.metric_collection_endpoint.is_none() {
-            // We don't start the consumption metrics task if this is not set in the config.
-            // So, no need to imitate the accesses in that case.
-            return;
-        }
-
-        // The consumption metrics are collected on a per-tenant basis, by a single
-        // global background loop.
-        // It limits the number of synthetic size calculations using the global
-        // `concurrent_tenant_size_logical_size_queries` semaphore to not overload
-        // the pageserver. (size calculation is somewhat expensive in terms of CPU and IOs).
-        //
-        // If we used that same semaphore here, then we'd compete for the
-        // same permits, which may impact timeliness of consumption metrics.
-        // That is a no-go, as consumption metrics are much more important
-        // than what we do here.
-        //
-        // So, we have a separate semaphore, initialized to the same
-        // number of permits as the `concurrent_tenant_size_logical_size_queries`.
-        // In the worst, we would have twice the amount of concurrenct size calculations.
-        // But in practice, the `p.threshold` >> `consumption metric interval`, and
-        // we spread out the eviction task using `random_init_delay`.
-        // So, the chance of the worst case is quite low in practice.
-        // It runs as a per-tenant task, but the eviction_task.rs is per-timeline.
-        // So, we must coordinate with other with other eviction tasks of this tenant.
-        let limit = self
-            .conf
-            .eviction_task_immitated_concurrent_logical_size_queries
-            .inner();
-
-        let mut throwaway_cache = HashMap::new();
-        let gather =
-            crate::tenant::size::gather_inputs(tenant, limit, None, &mut throwaway_cache, ctx);
-
-        tokio::select! {
-            _ = cancel.cancelled() => {}
-            gather_result = gather => {
-                match gather_result {
-                    Ok(_) => {},
-                    Err(e) => {
-                        // We don't care about the result, but, if it failed, we should log it,
-                        // since consumption metric might be hitting the cached value and
-                        // thus not encountering this error.
-                        warn!("failed to imitate synthetic size calculation accesses: {e:#}")
-                    }
-                }
-           }
-        }
-    }
 }

pageserver/src/tenant/timeline/walreceiver/walreceiver_connection.rs

@@ -2,7 +2,6 @@
 
 use std::{
     error::Error,
-    pin::pin,
     str::FromStr,
     sync::Arc,
     time::{Duration, SystemTime},
@@ -18,7 +17,7 @@ use postgres_ffi::v14::xlog_utils::normalize_lsn;
 use postgres_ffi::WAL_SEGMENT_SIZE;
 use postgres_protocol::message::backend::ReplicationMessage;
 use postgres_types::PgLsn;
-use tokio::{select, sync::watch, time};
+use tokio::{pin, select, sync::watch, time};
 use tokio_postgres::{replication::ReplicationStream, Client};
 use tokio_util::sync::CancellationToken;
 use tracing::{debug, error, info, trace, warn};
@@ -188,7 +187,8 @@ pub async fn handle_walreceiver_connection(
     let query = format!("START_REPLICATION PHYSICAL {startpoint}");
 
     let copy_stream = replication_client.copy_both_simple(&query).await?;
-    let mut physical_stream = pin!(ReplicationStream::new(copy_stream));
+    let physical_stream = ReplicationStream::new(copy_stream);
+    pin!(physical_stream);
 
     let mut waldecoder = WalStreamDecoder::new(startpoint, timeline.pg_version);
 

pgxn/neon/file_cache.c

@@ -14,7 +14,6 @@
  */
 
 #include <sys/file.h>
-#include <sys/statvfs.h>
 #include <unistd.h>
 #include <fcntl.h>
 
@@ -24,8 +23,6 @@
 #include "pgstat.h"
 #include "pagestore_client.h"
 #include "access/parallel.h"
-#include "common/hashfn.h"
-#include "lib/hyperloglog.h"
 #include "postmaster/bgworker.h"
 #include "storage/relfilenode.h"
 #include "storage/buf_internals.h"
@@ -37,9 +34,6 @@
 #include "storage/fd.h"
 #include "storage/pg_shmem.h"
 #include "storage/buf_internals.h"
-#include "storage/procsignal.h"
-#include "postmaster/bgworker.h"
-#include "postmaster/interrupt.h"
 
 /*
  * Local file cache is used to temporary store relations pages in local file system.
@@ -65,10 +59,6 @@
 
 #define SIZE_MB_TO_CHUNKS(size) ((uint32)((size) * MB / BLCKSZ / BLOCKS_PER_CHUNK))
 
-#define MAX_MONITOR_INTERVAL_USEC 1000000 /* 1 second */
-#define MAX_DISK_WRITE_RATE       1000 /* MB/sec */
-#define HYPER_LOG_LOG_BIT_WIDTH   10
-
 typedef struct FileCacheEntry
 {
 	BufferTag	key;
@@ -80,11 +70,8 @@ typedef struct FileCacheEntry
 
 typedef struct FileCacheControl
 {
-	uint32		size; /* size of cache file in chunks */
-	uint32		used; /* number of used chunks */
-	dlist_head	lru; /* double linked list for LRU replacement algorithm */
-	hyperLogLogState wss_estimation; /* estimation of wroking set size */
-	char hyperloglog_hashes[(1 << HYPER_LOG_LOG_BIT_WIDTH) + 1];
+	uint32 size; /* size of cache file in chunks */
+	dlist_head lru; /* double linked list for LRU replacement algorithm */
 } FileCacheControl;
 
 static HTAB* lfc_hash;
@@ -92,14 +79,12 @@ static int   lfc_desc;
 static LWLockId lfc_lock;
 static int   lfc_max_size;
 static int   lfc_size_limit;
-static int   lfc_free_space_watermark;
 static char* lfc_path;
 static  FileCacheControl* lfc_ctl;
 static shmem_startup_hook_type prev_shmem_startup_hook;
 #if PG_VERSION_NUM>=150000
 static shmem_request_hook_type prev_shmem_request_hook;
 #endif
-static int   lfc_shrinking_factor; /* power of two by which local cache size will be shrinked when lfc_free_space_watermark is reached */
 
 static void
 lfc_shmem_startup(void)
@@ -127,14 +112,7 @@ lfc_shmem_startup(void)
 								 &info,
 								 HASH_ELEM | HASH_BLOBS);
 		lfc_ctl->size = 0;
-		lfc_ctl->used = 0;
 		dlist_init(&lfc_ctl->lru);
-		initHyperLogLog(&lfc_ctl->wss_estimation, HYPER_LOG_LOG_BIT_WIDTH);
-
-		/* We need hashes in shared memory */
-		pfree(lfc_ctl->wss_estimation.hashesArr);
-		memset(lfc_ctl->hyperloglog_hashes, 0, sizeof lfc_ctl->hyperloglog_hashes);
-		lfc_ctl->wss_estimation.hashesArr = lfc_ctl->hyperloglog_hashes;
 
 		/* Remove file cache on restart */
 		(void)unlink(lfc_path);
@@ -187,7 +165,7 @@ lfc_change_limit_hook(int newval, void *extra)
 		}
 	}
 	LWLockAcquire(lfc_lock, LW_EXCLUSIVE);
-	while (new_size < lfc_ctl->used && !dlist_is_empty(&lfc_ctl->lru))
+	while (new_size < lfc_ctl->size && !dlist_is_empty(&lfc_ctl->lru))
 	{
 		/* Shrink cache by throwing away least recently accessed chunks and returning their space to file system */
 		FileCacheEntry* victim = dlist_container(FileCacheEntry, lru_node, dlist_pop_head_node(&lfc_ctl->lru));
@@ -197,86 +175,12 @@ lfc_change_limit_hook(int newval, void *extra)
 			elog(LOG, "Failed to punch hole in file: %m");
 #endif
 		hash_search(lfc_hash, &victim->key, HASH_REMOVE, NULL);
-		lfc_ctl->used -= 1;
+		lfc_ctl->size -= 1;
 	}
 	elog(LOG, "set local file cache limit to %d", new_size);
 	LWLockRelease(lfc_lock);
 }
 
-/*
- * Local file system state monitor check available free space.
- * If it is lower than lfc_free_space_watermark then we shrink size of local cache
- * but throwing away least recently accessed chunks.
- * First time low space watermark is reached cache size is divided by two,
- * second time by four,... Finally we remove all chunks from local cache.
- *
- * Please notice that we are not changing lfc_cache_size: it is used to be adjusted by autoscaler.
- * We only throw away cached chunks but do not prevent from filling cache by new chunks.
- *
- * Interval of poooling cache state is calculated as minimal time needed to consume lfc_free_space_watermark
- * disk space with maximal possible disk write speed (1Gb/sec). But not larger than 1 second.
- * Calling statvfs each second should not add any noticeable overhead.
- */
-void
-FileCacheMonitorMain(Datum main_arg)
-{
-	/*
-	 * Choose file system state monitor interval so that space can not be exosted
-	 * during this period but not longer than  MAX_MONITOR_INTERVAL (10 sec)
-	 */
-	uint64 monitor_interval = Min(MAX_MONITOR_INTERVAL_USEC, lfc_free_space_watermark*MB/MAX_DISK_WRITE_RATE);
-
-	/* Establish signal handlers. */
-	pqsignal(SIGUSR1, procsignal_sigusr1_handler);
-	pqsignal(SIGHUP, SignalHandlerForConfigReload);
-	pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
-	BackgroundWorkerUnblockSignals();
-
-	/* Periodically dump buffers until terminated. */
-	while (!ShutdownRequestPending)
-	{
-		if (lfc_size_limit != 0)
-		{
-			struct statvfs sfs;
-			if (statvfs(lfc_path, &sfs) < 0)
-			{
-				elog(WARNING, "Failed to obtain status of %s: %m", lfc_path);
-			}
-			else
-			{
-				if (sfs.f_bavail*sfs.f_bsize < lfc_free_space_watermark*MB)
-				{
-					if (lfc_shrinking_factor < 31) {
-						lfc_shrinking_factor += 1;
-					}
-					lfc_change_limit_hook(lfc_size_limit >> lfc_shrinking_factor, NULL);
-				}
-				else
-					lfc_shrinking_factor = 0; /* reset to initial value */
-			}
-		}
-		pg_usleep(monitor_interval);
-	}
-}
-
-static void
-lfc_register_free_space_monitor(void)
-{
-	BackgroundWorker bgw;
-	memset(&bgw, 0, sizeof(bgw));
-	bgw.bgw_flags = BGWORKER_SHMEM_ACCESS;
-	bgw.bgw_start_time = BgWorkerStart_RecoveryFinished;
-	snprintf(bgw.bgw_library_name, BGW_MAXLEN, "neon");
-	snprintf(bgw.bgw_function_name, BGW_MAXLEN, "FileCacheMonitorMain");
-	snprintf(bgw.bgw_name, BGW_MAXLEN, "Local free space monitor");
-	snprintf(bgw.bgw_type, BGW_MAXLEN, "Local free space monitor");
-	bgw.bgw_restart_time = 5;
-	bgw.bgw_notify_pid = 0;
-	bgw.bgw_main_arg = (Datum) 0;
-
-	RegisterBackgroundWorker(&bgw);
-}
-
 void
 lfc_init(void)
 {
@@ -313,19 +217,6 @@ lfc_init(void)
 							lfc_change_limit_hook,
 							NULL);
 
-	DefineCustomIntVariable("neon.free_space_watermark",
-							"Minimal free space in local file system after reaching which local file cache will be truncated",
-							NULL,
-							&lfc_free_space_watermark,
-							1024, /* 1GB */
-							0,
-							INT_MAX,
-							PGC_SIGHUP,
-							GUC_UNIT_MB,
-							NULL,
-							NULL,
-							NULL);
-
 	DefineCustomStringVariable("neon.file_cache_path",
 							   "Path to local file cache (can be raw device)",
 							   NULL,
@@ -340,9 +231,6 @@ lfc_init(void)
 	if (lfc_max_size == 0)
 		return;
 
-	if (lfc_free_space_watermark != 0)
-		lfc_register_free_space_monitor();
-
 	prev_shmem_startup_hook = shmem_startup_hook;
 	shmem_startup_hook = lfc_shmem_startup;
 #if PG_VERSION_NUM>=150000
@@ -407,11 +295,6 @@ lfc_read(RelFileNode rnode, ForkNumber forkNum, BlockNumber blkno,
 
 	LWLockAcquire(lfc_lock, LW_EXCLUSIVE);
 	entry = hash_search_with_hash_value(lfc_hash, &tag, hash, HASH_FIND, NULL);
-
-	/* Approximate working set */
-	tag.blockNum = blkno;
-	addHyperLogLog(&lfc_ctl->wss_estimation, hash_bytes((char const*)&tag, sizeof(tag)));
-
 	if (entry == NULL || (entry->bitmap[chunk_offs >> 5] & (1 << (chunk_offs & 31))) == 0)
 	{
 		/* Page is not cached */
@@ -497,7 +380,7 @@ lfc_write(RelFileNode rnode, ForkNumber forkNum, BlockNumber blkno,
 		 * there are should be very large number of concurrent IO operations and them are limited by max_connections,
 		 * we prefer not to complicate code and use second approach.
 		 */
-		if (lfc_ctl->used >= SIZE_MB_TO_CHUNKS(lfc_size_limit) && !dlist_is_empty(&lfc_ctl->lru))
+		if (lfc_ctl->size >= SIZE_MB_TO_CHUNKS(lfc_size_limit) && !dlist_is_empty(&lfc_ctl->lru))
 		{
 			/* Cache overflow: evict least recently used chunk */
 			FileCacheEntry* victim = dlist_container(FileCacheEntry, lru_node, dlist_pop_head_node(&lfc_ctl->lru));
@@ -507,10 +390,7 @@ lfc_write(RelFileNode rnode, ForkNumber forkNum, BlockNumber blkno,
 			elog(LOG, "Swap file cache page");
 		}
 		else
-		{
-			lfc_ctl->used += 1;
 			entry->offset = lfc_ctl->size++; /* allocate new chunk at end of file */
-		}
 		entry->access_count = 1;
 		memset(entry->bitmap, 0, sizeof entry->bitmap);
 	}
@@ -714,21 +594,3 @@ local_cache_pages(PG_FUNCTION_ARGS)
 	else
 		SRF_RETURN_DONE(funcctx);
 }
-
-PG_FUNCTION_INFO_V1(approximate_working_set_size);
-
-Datum
-approximate_working_set_size(PG_FUNCTION_ARGS)
-{
-	int32 dc = -1;
-	if (lfc_size_limit != 0)
-	{
-		bool reset = PG_GETARG_BOOL(0);
-		LWLockAcquire(lfc_lock, reset ? LW_EXCLUSIVE : LW_SHARED);
-		dc = (int32) estimateHyperLogLog(&lfc_ctl->wss_estimation);
-		if (reset)
-			memset(lfc_ctl->hyperloglog_hashes, 0, sizeof lfc_ctl->hyperloglog_hashes);
-		LWLockRelease(lfc_lock);
-	}
-	PG_RETURN_INT32(dc);
-}

pgxn/neon/neon--1.0.sql

@@ -27,13 +27,6 @@ RETURNS SETOF RECORD
 AS 'MODULE_PATHNAME', 'local_cache_pages'
 LANGUAGE C PARALLEL SAFE;
 
-CREATE FUNCTION approximate_working_set_size(reset bool)
-RETURNS integer
-AS 'MODULE_PATHNAME', 'approximate_working_set_size'
-LANGUAGE C PARALLEL SAFE;
-
-GRANT EXECUTE ON FUNCTION approximate_working_set_size() TO pg_monitor;
-
 -- Create a view for convenient access.
 CREATE VIEW local_cache AS
 	SELECT P.* FROM local_cache_pages() AS P

rust-toolchain.toml

@@ -1,5 +1,5 @@
 [toolchain]
-channel = "1.68.2"
+channel = "1.66.1"
 profile = "default"
 # The default profile includes rustc, rust-std, cargo, rust-docs, rustfmt and clippy.
 # https://rust-lang.github.io/rustup/concepts/profiles.html

safekeeper/src/bin/safekeeper.rs

@@ -5,7 +5,6 @@ use anyhow::{bail, Context, Result};
 use clap::Parser;
 use remote_storage::RemoteStorageConfig;
 use toml_edit::Document;
-use utils::signals::ShutdownSignals;
 
 use std::fs::{self, File};
 use std::io::{ErrorKind, Write};
@@ -40,7 +39,7 @@ use utils::{
     logging::{self, LogFormat},
     project_git_version,
     sentry_init::init_sentry,
-    tcp_listener,
+    signals, tcp_listener,
 };
 
 const PID_FILE_NAME: &str = "safekeeper.pid";
@@ -217,6 +216,7 @@ fn start_safekeeper(conf: SafeKeeperConf) -> Result<()> {
     let timeline_collector = safekeeper::metrics::TimelineCollector::new();
     metrics::register_internal(Box::new(timeline_collector))?;
 
+    let signals = signals::install_shutdown_handlers()?;
     let mut threads = vec![];
     let (wal_backup_launcher_tx, wal_backup_launcher_rx) = mpsc::channel(100);
 
@@ -274,12 +274,15 @@ fn start_safekeeper(conf: SafeKeeperConf) -> Result<()> {
 
     set_build_info_metric(GIT_VERSION);
     // TODO: put more thoughts into handling of failed threads
-    // We should catch & die if they are in trouble.
+    // We probably should restart them.
 
-    // On any shutdown signal, log receival and exit. Additionally, handling
-    // SIGQUIT prevents coredump.
-    ShutdownSignals::handle(|signal| {
-        info!("received {}, terminating", signal.name());
+    // NOTE: we still have to handle signals like SIGQUIT to prevent coredumps
+    signals.handle(|signal| {
+        // TODO: implement graceful shutdown with joining threads etc
+        info!(
+            "received {}, terminating in immediate shutdown mode",
+            signal.name()
+        );
         std::process::exit(0);
     })
 }

scripts/sk_collect_dumps/.gitignore

@@ -1,2 +0,0 @@
-result
-*.json

scripts/sk_collect_dumps/readme.md

@@ -1,25 +0,0 @@
-# Collect /v1/debug_dump from all safekeeper nodes
-
-1. Run ansible playbooks to collect .json dumps from all safekeepers and store them in `./result` directory.
-2. Run `DB_CONNSTR=... ./upload.sh prod_feb30` to upload dumps to `prod_feb30` table in specified postgres database.
-
-## How to use ansible (staging)
-
-```
-AWS_DEFAULT_PROFILE=dev ansible-playbook -i ../../.github/ansible/staging.us-east-2.hosts.yaml -e @../../.github/ansible/ssm_config remote.yaml
-
-AWS_DEFAULT_PROFILE=dev ansible-playbook -i ../../.github/ansible/staging.eu-west-1.hosts.yaml -e @../../.github/ansible/ssm_config remote.yaml
-```
-
-## How to use ansible (prod)
-
-```
-AWS_DEFAULT_PROFILE=prod ansible-playbook -i ../../.github/ansible/prod.us-west-2.hosts.yaml -e @../../.github/ansible/ssm_config remote.yaml
-
-AWS_DEFAULT_PROFILE=prod ansible-playbook -i ../../.github/ansible/prod.us-east-2.hosts.yaml -e @../../.github/ansible/ssm_config remote.yaml
-
-AWS_DEFAULT_PROFILE=prod ansible-playbook -i ../../.github/ansible/prod.eu-central-1.hosts.yaml -e @../../.github/ansible/ssm_config remote.yaml
-
-AWS_DEFAULT_PROFILE=prod ansible-playbook -i ../../.github/ansible/prod.ap-southeast-1.hosts.yaml -e @../../.github/ansible/ssm_config remote.yaml
-```
-

scripts/sk_collect_dumps/remote.yaml

@@ -1,18 +0,0 @@
-- name: Fetch state dumps from safekeepers
-  hosts: safekeepers
-  gather_facts: False
-  remote_user: "{{ remote_user }}"
-    
-  tasks:
-    - name: Download file
-      get_url:
-        url: "http://{{ inventory_hostname }}:7676/v1/debug_dump?dump_all=true&dump_disk_content=false"
-        dest: "/tmp/{{ inventory_hostname }}.json"
-
-    - name: Fetch file from remote hosts
-      fetch:
-        src: "/tmp/{{ inventory_hostname }}.json"
-        dest: "./result/{{ inventory_hostname }}.json"
-        flat: yes
-        fail_on_missing: no
-

scripts/sk_collect_dumps/upload.sh

@@ -1,52 +0,0 @@
-#!/bin/bash
-
-if [ -z "$DB_CONNSTR" ]; then
-    echo "DB_CONNSTR is not set"
-    exit 1
-fi
-
-# Create a temporary table for JSON data
-psql $DB_CONNSTR -c 'DROP TABLE IF EXISTS tmp_json'
-psql $DB_CONNSTR -c 'CREATE TABLE tmp_json (data jsonb)'
-
-for file in ./result/*.json; do
-    echo "$file"
-    SK_ID=$(jq '.config.id' $file)
-    echo "SK_ID: $SK_ID"
-    jq -c ".timelines[] |  . + {\"sk_id\": $SK_ID}" $file | psql $DB_CONNSTR -c "\\COPY tmp_json (data) FROM STDIN"
-done
-
-TABLE_NAME=$1
-
-if [ -z "$TABLE_NAME" ]; then
-    echo "TABLE_NAME is not set, skipping conversion to table with typed columns"
-    echo "Usage: ./upload.sh TABLE_NAME"
-    exit 0
-fi
-
-psql $DB_CONNSTR <<EOF
-CREATE TABLE $TABLE_NAME AS
-SELECT
-  (data->>'sk_id')::bigint AS sk_id,
-  (data->>'tenant_id') AS tenant_id,
-  (data->>'timeline_id') AS timeline_id,
-  (data->'memory'->>'active')::bool AS active,
-  (data->'memory'->>'flush_lsn')::bigint AS flush_lsn,
-  (data->'memory'->'mem_state'->>'backup_lsn')::bigint AS backup_lsn,
-  (data->'memory'->'mem_state'->>'commit_lsn')::bigint AS commit_lsn,
-  (data->'memory'->'mem_state'->>'peer_horizon_lsn')::bigint AS peer_horizon_lsn,
-  (data->'memory'->'mem_state'->>'remote_consistent_lsn')::bigint AS remote_consistent_lsn,
-  (data->'memory'->>'write_lsn')::bigint AS write_lsn,
-  (data->'memory'->>'num_computes')::bigint AS num_computes,
-  (data->'memory'->>'epoch_start_lsn')::bigint AS epoch_start_lsn,
-  (data->'memory'->>'last_removed_segno')::bigint AS last_removed_segno,
-  (data->'memory'->>'is_cancelled')::bool AS is_cancelled,
-  (data->'control_file'->>'backup_lsn')::bigint AS disk_backup_lsn,
-  (data->'control_file'->>'commit_lsn')::bigint AS disk_commit_lsn,
-  (data->'control_file'->'acceptor_state'->>'term')::bigint AS disk_term,
-  (data->'control_file'->>'local_start_lsn')::bigint AS local_start_lsn,
-  (data->'control_file'->>'peer_horizon_lsn')::bigint AS disk_peer_horizon_lsn,
-  (data->'control_file'->>'timeline_start_lsn')::bigint AS timeline_start_lsn,
-  (data->'control_file'->>'remote_consistent_lsn')::bigint AS disk_remote_consistent_lsn
-FROM tmp_json
-EOF

storage_broker/src/bin/storage_broker.rs

@@ -33,7 +33,6 @@ use tonic::transport::server::Connected;
 use tonic::Code;
 use tonic::{Request, Response, Status};
 use tracing::*;
-use utils::signals::ShutdownSignals;
 
 use metrics::{Encoder, TextEncoder};
 use storage_broker::metrics::{NUM_PUBS, NUM_SUBS_ALL, NUM_SUBS_TIMELINE};
@@ -438,14 +437,6 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
     info!("version: {GIT_VERSION}");
     ::metrics::set_build_info_metric(GIT_VERSION);
 
-    // On any shutdown signal, log receival and exit.
-    std::thread::spawn(move || {
-        ShutdownSignals::handle(|signal| {
-            info!("received {}, terminating", signal.name());
-            std::process::exit(0);
-        })
-    });
-
     let registry = Registry {
         shared_state: Arc::new(RwLock::new(SharedState::new(args.all_keys_chan_size))),
         timeline_chan_size: args.timeline_chan_size,

test_runner/fixtures/neon_fixtures.py

@@ -1220,6 +1220,28 @@ class PageserverHttpClient(requests.Session):
         self.verbose_error(res)
         return TenantConfig.from_json(res.json())
 
+    def set_tenant_config(self, tenant_id: TenantId, config: dict[str, Any]):
+        assert "tenant_id" not in config.keys()
+        res = self.put(
+            f"http://localhost:{self.port}/v1/tenant/config",
+            json={**config, "tenant_id": str(tenant_id)},
+        )
+        self.verbose_error(res)
+
+    def patch_tenant_config_client_side(
+        self,
+        tenant_id: TenantId,
+        inserts: Optional[Dict[str, Any]] = None,
+        removes: Optional[List[str]] = None,
+    ):
+        current = self.tenant_config(tenant_id).tenant_specific_overrides
+        if inserts is not None:
+            current.update(inserts)
+        if removes is not None:
+            for key in removes:
+                del current[key]
+        self.set_tenant_config(tenant_id, current)
+
     def tenant_size(self, tenant_id: TenantId) -> int:
         return self.tenant_size_and_modelinputs(tenant_id)[0]
 
@@ -1536,6 +1558,18 @@ class PageserverHttpClient(requests.Session):
         for layer in info.historic_layers:
             self.evict_layer(tenant_id, timeline_id, layer.layer_file_name)
 
+    def disk_usage_eviction_run(self, request: dict[str, Any]):
+        res = self.put(
+            f"http://localhost:{self.port}/v1/disk_usage_eviction/run",
+            json=request,
+        )
+        self.verbose_error(res)
+        return res.json()
+
+    def tenant_break(self, tenant_id: TenantId):
+        res = self.put(f"http://localhost:{self.port}/v1/tenant/{tenant_id}/break")
+        self.verbose_error(res)
+
 
 @dataclass
 class TenantConfig:

test_runner/regress/test_disk_usage_eviction.py

@@ -0,0 +1,541 @@
+import shutil
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Dict, Tuple
+
+import pytest
+import toml
+from fixtures.log_helper import log
+from fixtures.neon_fixtures import (
+    LocalFsStorage,
+    NeonEnv,
+    NeonEnvBuilder,
+    PageserverHttpClient,
+    PgBin,
+    RemoteStorageKind,
+    wait_for_last_flush_lsn,
+    wait_for_upload_queue_empty,
+    wait_until,
+)
+from fixtures.types import Lsn, TenantId, TimelineId
+
+GLOBAL_LRU_LOG_LINE = "tenant_min_resident_size-respecting LRU would not relieve pressure, evicting more following global LRU policy"
+
+
+@pytest.mark.parametrize("config_level_override", [None, 400])
+def test_min_resident_size_override_handling(
+    neon_env_builder: NeonEnvBuilder, config_level_override: int
+):
+    env = neon_env_builder.init_start()
+    ps_http = env.pageserver.http_client()
+
+    def assert_config(tenant_id, expect_override, expect_effective):
+        config = ps_http.tenant_config(tenant_id)
+        assert config.tenant_specific_overrides.get("min_resident_size_override") == expect_override
+        assert config.effective_config.get("min_resident_size_override") == expect_effective
+
+    def assert_overrides(tenant_id, default_tenant_conf_value):
+        ps_http.set_tenant_config(tenant_id, {"min_resident_size_override": 200})
+        assert_config(tenant_id, 200, 200)
+
+        ps_http.set_tenant_config(tenant_id, {"min_resident_size_override": 0})
+        assert_config(tenant_id, 0, 0)
+
+        ps_http.set_tenant_config(tenant_id, {})
+        assert_config(tenant_id, None, default_tenant_conf_value)
+
+    env.pageserver.stop()
+    if config_level_override is not None:
+        env.pageserver.start(
+            overrides=(
+                "--pageserver-config-override=tenant_config={ min_resident_size_override =  "
+                + str(config_level_override)
+                + " }",
+            )
+        )
+    else:
+        env.pageserver.start()
+
+    tenant_id, _ = env.neon_cli.create_tenant()
+    assert_overrides(tenant_id, config_level_override)
+
+    # Also ensure that specifying the paramter to create_tenant works, in addition to http-level recconfig.
+    tenant_id, _ = env.neon_cli.create_tenant(conf={"min_resident_size_override": "100"})
+    assert_config(tenant_id, 100, 100)
+    ps_http.set_tenant_config(tenant_id, {})
+    assert_config(tenant_id, None, config_level_override)
+
+
+@dataclass
+class EvictionEnv:
+    timelines: list[Tuple[TenantId, TimelineId]]
+    neon_env: NeonEnv
+    pg_bin: PgBin
+    pageserver_http: PageserverHttpClient
+    layer_size: int
+    pgbench_init_lsns: Dict[TenantId, Lsn]
+
+    def timelines_du(self) -> Tuple[int, int, int]:
+        return poor_mans_du(self.neon_env, [(tid, tlid) for tid, tlid in self.timelines])
+
+    def du_by_timeline(self) -> Dict[Tuple[TenantId, TimelineId], int]:
+        return {
+            (tid, tlid): poor_mans_du(self.neon_env, [(tid, tlid)])[0]
+            for tid, tlid in self.timelines
+        }
+
+    def warm_up_tenant(self, tenant_id: TenantId):
+        """
+        Start a read-only compute at the LSN after pgbench -i, and run pgbench -S against it.
+        This assumes that the tenant is still at the state after pbench -i.
+        """
+        lsn = self.pgbench_init_lsns[tenant_id]
+        with self.neon_env.postgres.create_start("main", tenant_id=tenant_id, lsn=lsn) as pg:
+            self.pg_bin.run(["pgbench", "-S", pg.connstr()])
+
+    def pageserver_start_with_disk_usage_eviction(
+        self, period, max_usage_pct, min_avail_bytes, mock_behavior
+    ):
+        disk_usage_config = {
+            "period": period,
+            "max_usage_pct": max_usage_pct,
+            "min_avail_bytes": min_avail_bytes,
+            "mock_statvfs": mock_behavior,
+        }
+
+        enc = toml.TomlEncoder()
+
+        self.neon_env.pageserver.start(
+            overrides=(
+                "--pageserver-config-override=disk_usage_based_eviction="
+                + enc.dump_inline_table(disk_usage_config).replace("\n", " "),
+            ),
+        )
+
+        def statvfs_called():
+            assert self.neon_env.pageserver.log_contains(".*running mocked statvfs.*")
+
+        wait_until(10, 1, statvfs_called)
+
+
+@pytest.fixture
+def eviction_env(request, neon_env_builder: NeonEnvBuilder, pg_bin: PgBin) -> EvictionEnv:
+    """
+    Creates two tenants, one somewhat larger than the other.
+    """
+
+    log.info(f"setting up eviction_env for test {request.node.name}")
+
+    neon_env_builder.enable_remote_storage(RemoteStorageKind.LOCAL_FS, f"{request.node.name}")
+
+    env = neon_env_builder.init_start()
+    pageserver_http = env.pageserver.http_client()
+
+    # allow because we are invoking this manually; we always warn on executing disk based eviction
+    env.pageserver.allowed_errors.append(r".* running disk usage based eviction due to pressure.*")
+
+    # remove the initial tenant
+    ## why wait for upload queue? => https://github.com/neondatabase/neon/issues/3865
+    assert env.initial_timeline
+    wait_for_upload_queue_empty(env.pageserver, env.initial_tenant, env.initial_timeline)
+    pageserver_http.tenant_detach(env.initial_tenant)
+    assert isinstance(env.remote_storage, LocalFsStorage)
+    tenant_remote_storage = env.remote_storage.root / "tenants" / str(env.initial_tenant)
+    assert tenant_remote_storage.is_dir()
+    shutil.rmtree(tenant_remote_storage)
+    env.initial_tenant = TenantId("0" * 32)
+    env.initial_timeline = None
+
+    # Choose small layer_size so that we can use low pgbench_scales and still get a large count of layers.
+    # Large count of layers and small layer size is good for testing because it makes evictions predictable.
+    # Predictable in the sense that many layer evictions will be required to reach the eviction target, because
+    # each eviction only makes small progress. That means little overshoot, and thereby stable asserts.
+    pgbench_scales = [4, 6]
+    layer_size = 5 * 1024**2
+
+    pgbench_init_lsns = {}
+
+    timelines = []
+    for scale in pgbench_scales:
+        tenant_id, timeline_id = env.neon_cli.create_tenant(
+            conf={
+                "gc_period": "0s",
+                "compaction_period": "0s",
+                "checkpoint_distance": f"{layer_size}",
+                "image_creation_threshold": "100",
+                "compaction_target_size": f"{layer_size}",
+            }
+        )
+
+        with env.postgres.create_start("main", tenant_id=tenant_id) as pg:
+            pg_bin.run(["pgbench", "-i", f"-s{scale}", pg.connstr()])
+            wait_for_last_flush_lsn(env, pg, tenant_id, timeline_id)
+
+        timelines.append((tenant_id, timeline_id))
+
+    # stop the safekeepers to avoid on-demand downloads caused by
+    # initial logical size calculation triggered by walreceiver connection status
+    # when we restart the pageserver process in any of the tests
+    env.neon_cli.safekeeper_stop()
+
+    # after stopping the safekeepers, we know that no new WAL will be coming in
+    for tenant_id, timeline_id in timelines:
+        pageserver_http.timeline_checkpoint(tenant_id, timeline_id)
+        wait_for_upload_queue_empty(env.pageserver, tenant_id, timeline_id)
+        tl_info = pageserver_http.timeline_detail(tenant_id, timeline_id)
+        assert tl_info["last_record_lsn"] == tl_info["disk_consistent_lsn"]
+        assert tl_info["disk_consistent_lsn"] == tl_info["remote_consistent_lsn"]
+        pgbench_init_lsns[tenant_id] = Lsn(tl_info["last_record_lsn"])
+
+        layers = pageserver_http.layer_map_info(tenant_id, timeline_id)
+        log.info(f"{layers}")
+        assert (
+            len(layers.historic_layers) >= 10
+        ), "evictions happen at layer granularity, but we often assert at byte-granularity"
+
+    eviction_env = EvictionEnv(
+        timelines=timelines,
+        neon_env=env,
+        pageserver_http=pageserver_http,
+        layer_size=layer_size,
+        pg_bin=pg_bin,
+        pgbench_init_lsns=pgbench_init_lsns,
+    )
+
+    return eviction_env
+
+
+def test_broken_tenants_are_skipped(eviction_env: EvictionEnv):
+    env = eviction_env
+
+    env.neon_env.pageserver.allowed_errors.append(
+        r".* Changing Active tenant to Broken state, reason: broken from test"
+    )
+    broken_tenant_id, broken_timeline_id = env.timelines[0]
+    env.pageserver_http.tenant_break(broken_tenant_id)
+
+    healthy_tenant_id, healthy_timeline_id = env.timelines[1]
+
+    broken_size_pre, _, _ = poor_mans_du(env.neon_env, [(broken_tenant_id, broken_timeline_id)])
+    healthy_size_pre, _, _ = poor_mans_du(env.neon_env, [(healthy_tenant_id, healthy_timeline_id)])
+
+    # try to evict everything, then validate that broken tenant wasn't touched
+    target = broken_size_pre + healthy_size_pre
+
+    response = env.pageserver_http.disk_usage_eviction_run({"evict_bytes": target})
+    log.info(f"{response}")
+
+    broken_size_post, _, _ = poor_mans_du(env.neon_env, [(broken_tenant_id, broken_timeline_id)])
+    healthy_size_post, _, _ = poor_mans_du(env.neon_env, [(healthy_tenant_id, healthy_timeline_id)])
+
+    assert broken_size_pre == broken_size_post, "broken tenant should not be touched"
+    assert healthy_size_post < healthy_size_pre
+    assert healthy_size_post == 0
+    env.neon_env.pageserver.allowed_errors.append(".*" + GLOBAL_LRU_LOG_LINE)
+
+
+def test_pageserver_evicts_until_pressure_is_relieved(eviction_env: EvictionEnv):
+    """
+    Basic test to ensure that we evict enough to relieve pressure.
+    """
+    env = eviction_env
+    pageserver_http = env.pageserver_http
+
+    (total_on_disk, _, _) = env.timelines_du()
+
+    target = total_on_disk // 2
+
+    response = pageserver_http.disk_usage_eviction_run({"evict_bytes": target})
+    log.info(f"{response}")
+
+    (later_total_on_disk, _, _) = env.timelines_du()
+
+    actual_change = total_on_disk - later_total_on_disk
+
+    assert 0 <= actual_change, "nothing can load layers during this test"
+    assert actual_change >= target, "must evict more than half"
+    assert (
+        response["Finished"]["assumed"]["projected_after"]["freed_bytes"] >= actual_change
+    ), "report accurately evicted bytes"
+    assert response["Finished"]["assumed"]["failed"]["count"] == 0, "zero failures expected"
+
+
+def test_pageserver_respects_overridden_resident_size(eviction_env: EvictionEnv):
+    """
+    Override tenant min resident and ensure that it will be respected by eviction.
+    """
+    env = eviction_env
+    ps_http = env.pageserver_http
+
+    (total_on_disk, _, _) = env.timelines_du()
+    du_by_timeline = env.du_by_timeline()
+    log.info("du_by_timeline: %s", du_by_timeline)
+
+    assert len(du_by_timeline) == 2, "this test assumes two tenants"
+    large_tenant = max(du_by_timeline, key=du_by_timeline.__getitem__)
+    small_tenant = min(du_by_timeline, key=du_by_timeline.__getitem__)
+    assert du_by_timeline[large_tenant] > du_by_timeline[small_tenant]
+    assert (
+        du_by_timeline[large_tenant] - du_by_timeline[small_tenant] > 5 * env.layer_size
+    ), "ensure this test will do more than 1 eviction"
+
+    # Give the larger tenant a haircut while preventing the smaller tenant from getting one.
+    # To prevent the smaller from getting a haircut, we set min_resident_size to its current size.
+    # To ensure the larger tenant is getting a haircut, any non-zero `target` will do.
+    min_resident_size = du_by_timeline[small_tenant]
+    target = 1
+    assert (
+        du_by_timeline[large_tenant] > min_resident_size
+    ), "ensure the larger tenant will get a haircut"
+    ps_http.patch_tenant_config_client_side(
+        small_tenant[0], {"min_resident_size_override": min_resident_size}
+    )
+    ps_http.patch_tenant_config_client_side(
+        large_tenant[0], {"min_resident_size_override": min_resident_size}
+    )
+
+    # Make the large tenant more-recently used. An incorrect implemention would try to evict
+    # the smaller tenant completely first, before turning to the larger tenant,
+    # since the smaller tenant's layers are least-recently-used.
+    env.warm_up_tenant(large_tenant[0])
+
+    # do one run
+    response = ps_http.disk_usage_eviction_run({"evict_bytes": target})
+    log.info(f"{response}")
+
+    time.sleep(1)  # give log time to flush
+    assert not env.neon_env.pageserver.log_contains(
+        GLOBAL_LRU_LOG_LINE,
+    ), "this test is pointless if it fell back to global LRU"
+
+    (later_total_on_disk, _, _) = env.timelines_du()
+    later_du_by_timeline = env.du_by_timeline()
+    log.info("later_du_by_timeline: %s", later_du_by_timeline)
+
+    actual_change = total_on_disk - later_total_on_disk
+    assert 0 <= actual_change, "nothing can load layers during this test"
+    assert actual_change >= target, "eviction must always evict more than target"
+    assert (
+        response["Finished"]["assumed"]["projected_after"]["freed_bytes"] >= actual_change
+    ), "report accurately evicted bytes"
+    assert response["Finished"]["assumed"]["failed"]["count"] == 0, "zero failures expected"
+
+    assert (
+        later_du_by_timeline[small_tenant] == du_by_timeline[small_tenant]
+    ), "small tenant sees no haircut"
+    assert (
+        later_du_by_timeline[large_tenant] < du_by_timeline[large_tenant]
+    ), "large tenant gets a haircut"
+    assert du_by_timeline[large_tenant] - later_du_by_timeline[large_tenant] >= target
+
+
+def test_pageserver_falls_back_to_global_lru(eviction_env: EvictionEnv):
+    """
+    If we can't relieve pressure using tenant_min_resident_size-respecting eviction,
+    we should continue to evict layers following global LRU.
+    """
+    env = eviction_env
+    ps_http = env.pageserver_http
+
+    (total_on_disk, _, _) = env.timelines_du()
+    target = total_on_disk
+
+    response = ps_http.disk_usage_eviction_run({"evict_bytes": target})
+    log.info(f"{response}")
+
+    (later_total_on_disk, _, _) = env.timelines_du()
+    actual_change = total_on_disk - later_total_on_disk
+    assert 0 <= actual_change, "nothing can load layers during this test"
+    assert actual_change >= target, "eviction must always evict more than target"
+
+    time.sleep(1)  # give log time to flush
+    assert env.neon_env.pageserver.log_contains(GLOBAL_LRU_LOG_LINE)
+    env.neon_env.pageserver.allowed_errors.append(".*" + GLOBAL_LRU_LOG_LINE)
+
+
+def test_partial_evict_tenant(eviction_env: EvictionEnv):
+    """
+    Warm up a tenant, then build up pressure to cause in evictions in both.
+    We expect
+    * the default min resident size to be respect (largest layer file size)
+    * the warmed-up tenants layers above min resident size to be evicted after the cold tenant's.
+    """
+    env = eviction_env
+    ps_http = env.pageserver_http
+
+    (total_on_disk, _, _) = env.timelines_du()
+    du_by_timeline = env.du_by_timeline()
+
+    # pick any tenant
+    [our_tenant, other_tenant] = list(du_by_timeline.keys())
+    (tenant_id, timeline_id) = our_tenant
+
+    # make our tenant more recently used than the other one
+    env.warm_up_tenant(tenant_id)
+
+    # Build up enough pressure to require evictions from both tenants,
+    # but not enough to fall into global LRU.
+    # So, set target to all occipied space, except 2*env.layer_size per tenant
+    target = (
+        du_by_timeline[other_tenant] + (du_by_timeline[our_tenant] // 2) - 2 * 2 * env.layer_size
+    )
+    response = ps_http.disk_usage_eviction_run({"evict_bytes": target})
+    log.info(f"{response}")
+
+    (later_total_on_disk, _, _) = env.timelines_du()
+    actual_change = total_on_disk - later_total_on_disk
+    assert 0 <= actual_change, "nothing can load layers during this test"
+    assert actual_change >= target, "eviction must always evict more than target"
+
+    later_du_by_timeline = env.du_by_timeline()
+    for tenant, later_tenant_usage in later_du_by_timeline.items():
+        assert (
+            later_tenant_usage < du_by_timeline[tenant]
+        ), "all tenants should have lost some layers"
+
+    assert (
+        later_du_by_timeline[our_tenant] > 0.5 * du_by_timeline[our_tenant]
+    ), "our warmed up tenant should be at about half capacity, part 1"
+    assert (
+        # We don't know exactly whether the cold tenant needs 2 or just 1 env.layer_size wiggle room.
+        # So, check for up to 3 here.
+        later_du_by_timeline[our_tenant]
+        < 0.5 * du_by_timeline[our_tenant] + 3 * env.layer_size
+    ), "our warmed up tenant should be at about half capacity, part 2"
+    assert (
+        later_du_by_timeline[other_tenant] < 2 * env.layer_size
+    ), "the other tenant should be evicted to is min_resident_size, i.e., max layer file size"
+
+
+def poor_mans_du(
+    env: NeonEnv, timelines: list[Tuple[TenantId, TimelineId]]
+) -> Tuple[int, int, int]:
+    """
+    Disk usage, largest, smallest layer for layer files over the given (tenant, timeline) tuples;
+    this could be done over layers endpoint just as well.
+    """
+    total_on_disk = 0
+    largest_layer = 0
+    smallest_layer = None
+    for tenant_id, timeline_id in timelines:
+        dir = Path(env.repo_dir) / "tenants" / str(tenant_id) / "timelines" / str(timeline_id)
+        assert dir.exists(), f"timeline dir does not exist: {dir}"
+        sum = 0
+        for file in dir.iterdir():
+            if "__" not in file.name:
+                continue
+            size = file.stat().st_size
+            sum += size
+            largest_layer = max(largest_layer, size)
+            if smallest_layer:
+                smallest_layer = min(smallest_layer, size)
+            else:
+                smallest_layer = size
+            log.info(f"{tenant_id}/{timeline_id} => {file.name} {size}")
+
+        log.info(f"{tenant_id}/{timeline_id}: sum {sum}")
+        total_on_disk += sum
+
+    assert smallest_layer is not None or total_on_disk == 0 and largest_layer == 0
+    return (total_on_disk, largest_layer, smallest_layer or 0)
+
+
+def test_statvfs_error_handling(eviction_env: EvictionEnv):
+    """
+    We should log an error that statvfs fails.
+    """
+    env = eviction_env
+    env.neon_env.pageserver.stop()
+    env.pageserver_start_with_disk_usage_eviction(
+        period="1s",
+        max_usage_pct=90,
+        min_avail_bytes=0,
+        mock_behavior={
+            "type": "Failure",
+            "mocked_error": "EIO",
+        },
+    )
+
+    assert env.neon_env.pageserver.log_contains(".*statvfs failed.*EIO")
+    env.neon_env.pageserver.allowed_errors.append(".*statvfs failed.*EIO")
+
+
+def test_statvfs_pressure_usage(eviction_env: EvictionEnv):
+    """
+    If statvfs data shows 100% usage, the eviction task will drive it down to
+    the configured max_usage_pct.
+    """
+    env = eviction_env
+
+    env.neon_env.pageserver.stop()
+
+    # make it seem like we're at 100% utilization by setting total bytes to the used bytes
+    total_size, _, _ = env.timelines_du()
+    blocksize = 512
+    total_blocks = (total_size + (blocksize - 1)) // blocksize
+
+    env.pageserver_start_with_disk_usage_eviction(
+        period="1s",
+        max_usage_pct=33,
+        min_avail_bytes=0,
+        mock_behavior={
+            "type": "Success",
+            "blocksize": blocksize,
+            "total_blocks": total_blocks,
+            # Only count layer files towards used bytes in the mock_statvfs.
+            # This avoids accounting for metadata files & tenant conf in the tests.
+            "name_filter": ".*__.*",
+        },
+    )
+
+    def relieved_log_message():
+        assert env.neon_env.pageserver.log_contains(".*disk usage pressure relieved")
+
+    wait_until(10, 1, relieved_log_message)
+
+    post_eviction_total_size, _, _ = env.timelines_du()
+
+    assert post_eviction_total_size <= 0.33 * total_size, "we requested max 33% usage"
+
+
+def test_statvfs_pressure_min_avail_bytes(eviction_env: EvictionEnv):
+    """
+    If statvfs data shows 100% usage, the eviction task will drive it down to
+    at least the configured min_avail_bytes.
+    """
+    env = eviction_env
+
+    env.neon_env.pageserver.stop()
+
+    # make it seem like we're at 100% utilization by setting total bytes to the used bytes
+    total_size, _, _ = env.timelines_du()
+    blocksize = 512
+    total_blocks = (total_size + (blocksize - 1)) // blocksize
+
+    min_avail_bytes = total_size // 3
+
+    env.pageserver_start_with_disk_usage_eviction(
+        period="1s",
+        max_usage_pct=100,
+        min_avail_bytes=min_avail_bytes,
+        mock_behavior={
+            "type": "Success",
+            "blocksize": blocksize,
+            "total_blocks": total_blocks,
+            # Only count layer files towards used bytes in the mock_statvfs.
+            # This avoids accounting for metadata files & tenant conf in the tests.
+            "name_filter": ".*__.*",
+        },
+    )
+
+    def relieved_log_message():
+        assert env.neon_env.pageserver.log_contains(".*disk usage pressure relieved")
+
+    wait_until(10, 1, relieved_log_message)
+
+    post_eviction_total_size, _, _ = env.timelines_du()
+
+    assert (
+        total_size - post_eviction_total_size >= min_avail_bytes
+    ), "we requested at least min_avail_bytes worth of free space"

test_runner/regress/test_timeline_delete.py

@@ -10,7 +10,7 @@ def test_timeline_delete(neon_simple_env: NeonEnv):
     env.pageserver.allowed_errors.append(".*Timeline .* was not found.*")
     env.pageserver.allowed_errors.append(".*timeline not found.*")
     env.pageserver.allowed_errors.append(".*Cannot delete timeline which has child timelines.*")
-    env.pageserver.allowed_errors.append(".*Precondition failed: Requested tenant is missing.*")
+    env.pageserver.allowed_errors.append(".*NotFound: tenant .*")
 
     ps_http = env.pageserver.http_client()
 
@@ -24,11 +24,11 @@ def test_timeline_delete(neon_simple_env: NeonEnv):
     invalid_tenant_id = TenantId.generate()
     with pytest.raises(
         PageserverApiException,
-        match="Precondition failed: Requested tenant is missing",
+        match=f"NotFound: tenant {invalid_tenant_id}",
     ) as exc:
         ps_http.timeline_delete(tenant_id=invalid_tenant_id, timeline_id=invalid_timeline_id)
 
-    assert exc.value.status_code == 412
+    assert exc.value.status_code == 404
 
     # construct pair of branches to validate that pageserver prohibits
     # deletion of ancestor timelines when they have child branches