perf(relay): cache Git pack hydration (#2169)

.env.example

@@ -66,12 +66,15 @@ RELAY_URL=ws://localhost:3000
 # -----------------------------------------------------------------------------
 # Git (NIP-34 bare repositories)
 # -----------------------------------------------------------------------------
-# Root directory for bare git repos. Repos are stored at
+# Root directory for ephemeral Git workspaces and the disposable pack cache.
-# {path}/{owner_hex}/{repo_id}.git/. Default: ./repos (relative to CWD).
+# Default: ./repos (relative to CWD).
-# Set an absolute path to keep repos stable across worktrees.
 # BUZZ_GIT_REPO_PATH=./repos
 # BUZZ_GIT_MAX_PACK_BYTES=524288000
 # BUZZ_GIT_MAX_REPO_BYTES=1048576000
+# Process-local immutable pack/index cache. Zero disables retention.
+# BUZZ_GIT_PACK_CACHE_PATH=./repos/.pack-cache
+# BUZZ_GIT_PACK_CACHE_MAX_BYTES=5368709120
+# BUZZ_GIT_PACK_CACHE_MAX_CONCURRENT_POPULATIONS=2
 
 # -----------------------------------------------------------------------------
 # Media Upload Admission

crates/buzz-relay/src/api/git/hydrate.rs

@@ -6,11 +6,11 @@
 //!
 //! 1. GET pointer → manifest digest.
 //! 2. GET manifest (digest-verified) → parsed [`Manifest`].
-//! 3. GET every pack the manifest names (digest-verified, in parallel).
+//! 3. Resolve every pack through the bounded local content-addressed cache.
-//! 4. **Phase 1** — for each pack: write `pack-<hex>.pack`; install and
+//! 4. **Phase 1** — hard-link or copy each verified pack/index pair into the
-//!    validate a cached `idx/<hex>` sidecar when present, otherwise run
+//!    ephemeral repo. A cache miss reads object storage and validates or
-//!    `git index-pack` to materialize `.idx`. Failure here tears down the
+//!    generates the index once per digest. Failure here tears down the tempdir
-//!    tempdir with no refs/HEAD ever written.
+//!    with no refs/HEAD ever written.
 //! 5. **Phase 2** — only after all packs are indexed: write loose refs and
 //!    HEAD from the manifest.
 //!
@@ -19,8 +19,8 @@
 //! mid-stream. Sami/Max named this explicitly.
 //!
 //! The returned [`HydratedRepo`] owns a [`tempfile::TempDir`]; dropping it
-//! cleans up. Every read currently re-hydrates from scratch — naive but
+//! cleans up. Cached pack/index pairs are immutable performance state; object
-//! correct; caching is named follow-up work.
+//! storage remains authoritative.
 
 // Public surface is consumed by `transport.rs` after Eva's `AppState::git_store`
 // wires in; the items below are intentionally `pub` to keep the consumer-side
@@ -35,6 +35,7 @@ use tokio::process::Command;
 
 use super::cas_publish::ParentState;
 use super::manifest::{is_hex_oid, is_safe_refname, pointer_key, Manifest, ManifestError};
+use super::pack_cache::GitPackCache;
 use super::store::{ETag, GitStore, StoreError};
 use buzz_core::TenantContext;
 
@@ -75,6 +76,19 @@ impl HydratedRepo {
     }
 }
 
+/// Process-local resources and limits used to materialize one repository.
+#[derive(Clone, Copy)]
+pub struct HydrationOptions<'a> {
+    /// Immutable pack/index cache shared by Git operations in this process.
+    pub pack_cache: &'a GitPackCache,
+    /// Parent directory for ephemeral bare repositories.
+    pub scratch_dir: &'a Path,
+    /// Maximum bytes accepted for one pack object.
+    pub max_pack_bytes: u64,
+    /// Maximum aggregate pack bytes accepted for one repository.
+    pub max_repo_bytes: u64,
+}
+
 /// Hydration errors.
 ///
 /// "Repo doesn't exist" is signalled by `Ok(None)` from `hydrate_for_read`,
@@ -112,21 +126,10 @@ pub async fn hydrate_for_read(
     ctx: &TenantContext,
     owner: &str,
     repo: &str,
-    scratch_dir: &Path,
+    options: HydrationOptions<'_>,
-    max_pack_bytes: u64,
-    max_repo_bytes: u64,
 ) -> Result<Option<HydratedRepo>, HydrateError> {
     let started_at = std::time::Instant::now();
-    let result = hydrate_for_read_inner(
+    let result = hydrate_for_read_inner(store, ctx, owner, repo, options).await;
-        store,
-        ctx,
-        owner,
-        repo,
-        scratch_dir,
-        max_pack_bytes,
-        max_repo_bytes,
-    )
-    .await;
     let outcome = match &result {
         Ok(Some(repo)) => {
             metrics::histogram!("buzz_git_hydrate_bytes").record(repo.hydrated_bytes() as f64);
@@ -151,23 +154,12 @@ async fn hydrate_for_read_inner(
     ctx: &TenantContext,
     owner: &str,
     repo: &str,
-    scratch_dir: &Path,
+    options: HydrationOptions<'_>,
-    max_pack_bytes: u64,
-    max_repo_bytes: u64,
 ) -> Result<Option<HydratedRepo>, HydrateError> {
     let Some((_etag, _digest, manifest)) = load_pointer(store, ctx, owner, repo).await? else {
         return Ok(None);
     };
-    Ok(Some(
+    Ok(Some(materialize_manifest(store, &manifest, options).await?))
-        materialize_manifest(
-            store,
-            &manifest,
-            scratch_dir,
-            max_pack_bytes,
-            max_repo_bytes,
-        )
-        .await?,
-    ))
 }
 
 /// Load the current manifest for a read path without materializing pack objects.
@@ -212,20 +204,11 @@ pub async fn hydrate_for_write(
     ctx: &TenantContext,
     owner: &str,
     repo: &str,
-    scratch_dir: &Path,
+    options: HydrationOptions<'_>,
-    max_pack_bytes: u64,
-    max_repo_bytes: u64,
 ) -> Result<(HydratedRepo, ParentState), HydrateError> {
     match load_pointer(store, ctx, owner, repo).await? {
         Some((etag, digest, manifest)) => {
-            let repo = materialize_manifest(
+            let repo = materialize_manifest(store, &manifest, options).await?;
-                store,
-                &manifest,
-                scratch_dir,
-                max_pack_bytes,
-                max_repo_bytes,
-            )
-            .await?;
             let parent = ParentState::from_loaded(etag, digest, manifest);
             Ok((repo, parent))
         }
@@ -233,8 +216,9 @@ pub async fn hydrate_for_write(
             // First push: empty bare repo. No packs to fetch, no refs to
             // install. `receive-pack` will accept whatever the client
             // sends; `cas_publish` will use `If-None-Match: *`.
-            let tempdir = TempDir::new_in(scratch_dir)
+            let tempdir = TempDir::new_in(options.scratch_dir).map_err(|e| {
-                .map_err(|e| HydrateError::Hydrate(format!("tempdir in {scratch_dir:?}: {e}")))?;
+                HydrateError::Hydrate(format!("tempdir in {:?}: {e}", options.scratch_dir))
+            })?;
             let path = tempdir.path().to_path_buf();
             run_git(&path, &["init", "--bare", "--quiet"]).await?;
             Ok((
@@ -280,7 +264,7 @@ async fn load_pointer(
     Ok(Some((etag, digest, manifest)))
 }
 
-async fn get_verified_limited(
+pub(super) async fn get_verified_limited(
     store: &GitStore,
     key: &str,
     digest: &str,
@@ -304,13 +288,11 @@ async fn get_verified_limited(
 async fn materialize_manifest(
     store: &GitStore,
     manifest: &Manifest,
-    scratch_dir: &Path,
+    options: HydrationOptions<'_>,
-    max_pack_bytes: u64,
-    max_repo_bytes: u64,
 ) -> Result<HydratedRepo, HydrateError> {
     // Init bare repo.
-    let tempdir = TempDir::new_in(scratch_dir)
+    let tempdir = TempDir::new_in(options.scratch_dir)
-        .map_err(|e| HydrateError::Hydrate(format!("tempdir in {scratch_dir:?}: {e}")))?;
+        .map_err(|e| HydrateError::Hydrate(format!("tempdir in {:?}: {e}", options.scratch_dir)))?;
     let path = tempdir.path().to_path_buf();
     run_git(&path, &["init", "--bare", "--quiet"]).await?;
 
@@ -324,27 +306,25 @@ async fn materialize_manifest(
         let digest = key
             .strip_prefix("packs/")
             .ok_or_else(|| HydrateError::Hydrate(format!("malformed pack key {key:?}")))?;
-        let bytes = get_verified_limited(store, key, digest, max_pack_bytes).await?;
+        let pack_bytes = options
-        let pack_bytes = u64::try_from(bytes.len()).unwrap_or(u64::MAX);
+            .pack_cache
-        if pack_bytes > max_pack_bytes {
+            .materialize_pack(store, key, digest, &pack_dir, options.max_pack_bytes)
+            .await?;
+        if pack_bytes > options.max_pack_bytes {
             return Err(HydrateError::ResourceLimit(format!(
-                "pack {digest} is {pack_bytes} bytes (max {max_pack_bytes})"
+                "pack {digest} is {pack_bytes} bytes (max {})",
+                options.max_pack_bytes
             )));
         }
         hydrated_bytes = hydrated_bytes.checked_add(pack_bytes).ok_or_else(|| {
             HydrateError::ResourceLimit("repo byte count overflowed u64".to_string())
         })?;
-        if hydrated_bytes > max_repo_bytes {
+        if hydrated_bytes > options.max_repo_bytes {
             return Err(HydrateError::ResourceLimit(format!(
-                "repo needs {hydrated_bytes} hydrated bytes (max {max_repo_bytes})"
+                "repo needs {hydrated_bytes} hydrated bytes (max {})",
+                options.max_repo_bytes
             )));
         }
-
-        let pack_path = pack_dir.join(format!("pack-{digest}.pack"));
-        tokio::fs::write(&pack_path, &bytes)
-            .await
-            .map_err(|e| HydrateError::Hydrate(format!("write pack {digest}: {e}")))?;
-        install_or_generate_idx(store, &path, digest, &pack_path).await?;
     }
 
     // Phase 2: install refs and HEAD. After this point, the repo advertises.
@@ -393,14 +373,15 @@ async fn materialize_manifest(
     })
 }
 
-async fn install_or_generate_idx(
+pub(super) async fn install_or_generate_idx(
     store: &GitStore,
     repo_path: &Path,
     pack_digest: &str,
     pack_path: &Path,
+    max_idx_bytes: u64,
 ) -> Result<(), HydrateError> {
     let idx_path = pack_path.with_extension("idx");
-    match store.get_idx(pack_digest).await {
+    match store.get_idx(pack_digest, max_idx_bytes).await {
         Ok(Some(idx_bytes)) => {
             tokio::fs::write(&idx_path, &idx_bytes)
                 .await
@@ -438,7 +419,27 @@ async fn install_or_generate_idx(
     let pack_path_str = pack_path
         .to_str()
         .ok_or_else(|| HydrateError::Hydrate("pack path is not valid utf-8".to_string()))?;
-    run_git(repo_path, &["index-pack", pack_path_str]).await
+    run_git(repo_path, &["index-pack", pack_path_str]).await?;
+    enforce_idx_size(&idx_path, pack_digest, max_idx_bytes).await
+}
+
+async fn enforce_idx_size(
+    idx_path: &Path,
+    pack_digest: &str,
+    max_idx_bytes: u64,
+) -> Result<(), HydrateError> {
+    let idx_bytes = tokio::fs::metadata(&idx_path)
+        .await
+        .map_err(|error| {
+            HydrateError::Hydrate(format!("stat generated idx {pack_digest}: {error}"))
+        })?
+        .len();
+    if idx_bytes > max_idx_bytes {
+        return Err(HydrateError::ResourceLimit(format!(
+            "generated idx {pack_digest} is {idx_bytes} bytes (max {max_idx_bytes})"
+        )));
+    }
+    Ok(())
 }
 
 /// Run `git <args>` in `cwd`, fail on non-zero exit.
@@ -499,6 +500,17 @@ mod tests {
         assert!(!is_hex_oid(""));
     }
 
+    #[tokio::test]
+    async fn generated_idx_size_is_bounded() {
+        let scratch = TempDir::new().expect("scratch");
+        let idx_path = scratch.path().join("pack-test.idx");
+        tokio::fs::write(&idx_path, [0u8; 2]).await.expect("idx");
+
+        let result = enforce_idx_size(&idx_path, "test", 1).await;
+
+        assert!(matches!(result, Err(HydrateError::ResourceLimit(_))));
+    }
+
     // `pointer_key` is tested in `super::manifest::tests` — single source.
 
     fn tenant() -> TenantContext {
@@ -520,10 +532,20 @@ mod tests {
             packs: Vec::new(),
             parent: None,
         };
+        let cache = GitPackCache::new(scratch.path(), u64::MAX, 2).expect("cache");
 
-        let hydrated = materialize_manifest(&store, &manifest, scratch.path(), u64::MAX, u64::MAX)
+        let hydrated = materialize_manifest(
-            .await
+            &store,
-            .expect("materialize empty repo");
+            &manifest,
+            HydrationOptions {
+                pack_cache: &cache,
+                scratch_dir: scratch.path(),
+                max_pack_bytes: u64::MAX,
+                max_repo_bytes: u64::MAX,
+            },
+        )
+        .await
+        .expect("materialize empty repo");
 
         assert!(hydrated.path().starts_with(scratch.path()));
     }
@@ -643,11 +665,22 @@ mod tests {
 
         // Hydrate.
         let scratch = TempDir::new().unwrap();
-        let hydrated =
+        let cache = GitPackCache::new(scratch.path(), u64::MAX, 2).expect("cache");
-            hydrate_for_read(&st, &ctx, &owner, repo, scratch.path(), u64::MAX, u64::MAX)
+        let hydrated = hydrate_for_read(
-                .await
+            &st,
-                .expect("hydrate")
+            &ctx,
-                .expect("hydrate Some");
+            &owner,
+            repo,
+            HydrationOptions {
+                pack_cache: &cache,
+                scratch_dir: scratch.path(),
+                max_pack_bytes: u64::MAX,
+                max_repo_bytes: u64::MAX,
+            },
+        )
+        .await
+        .expect("hydrate")
+        .expect("hydrate Some");
         eprintln!("hydrated to {}", hydrated.path().display());
 
         // The hydrated repo must list the same ref with the same oid.
@@ -718,14 +751,18 @@ mod tests {
         let owner = format!("nope-{}", uuid::Uuid::new_v4());
         let ctx = tenant();
         let scratch = TempDir::new().unwrap();
+        let cache = GitPackCache::new(scratch.path(), u64::MAX, 2).expect("cache");
         let result = hydrate_for_read(
             &st,
             &ctx,
             &owner,
             "ghost",
-            scratch.path(),
+            HydrationOptions {
-            u64::MAX,
+                pack_cache: &cache,
-            u64::MAX,
+                scratch_dir: scratch.path(),
+                max_pack_bytes: u64::MAX,
+                max_repo_bytes: u64::MAX,
+            },
         )
         .await
         .expect("ok");
@@ -774,14 +811,18 @@ mod tests {
         }
 
         let scratch = TempDir::new().unwrap();
+        let cache = GitPackCache::new(scratch.path(), u64::MAX, 2).expect("cache");
         let hydrated = hydrate_for_read(
             &st,
             &ctx,
             &owner,
             "void",
-            scratch.path(),
+            HydrationOptions {
-            u64::MAX,
+                pack_cache: &cache,
-            u64::MAX,
+                scratch_dir: scratch.path(),
+                max_pack_bytes: u64::MAX,
+                max_repo_bytes: u64::MAX,
+            },
         )
         .await
         .expect("hydrate")

crates/buzz-relay/src/api/git/mod.rs

@@ -27,6 +27,7 @@ pub mod hook;
 pub mod hydrate;
 pub mod manifest;
 pub mod manifest_event;
+pub mod pack_cache;
 pub mod policy;
 pub mod store;
 pub mod transport;

crates/buzz-relay/src/api/git/pack_cache.rs

@@ -0,0 +1,686 @@
+//! Bounded local cache for immutable Git pack/index pairs.
+//!
+//! Object storage remains the durable source of truth. Cache entries are
+//! content-addressed by the verified pack digest and are published by an
+//! atomic directory rename only after both the pack and index are ready.
+
+use std::collections::HashMap;
+use std::path::{Path, PathBuf};
+use std::sync::atomic::{AtomicUsize, Ordering};
+use std::sync::{Arc, Mutex};
+use std::time::{Duration, Instant, SystemTime};
+
+use dashmap::mapref::entry::Entry;
+use dashmap::DashMap;
+use tempfile::Builder;
+
+use super::hydrate::{get_verified_limited, install_or_generate_idx, HydrateError};
+use super::store::GitStore;
+
+const HEARTBEAT_INTERVAL: Duration = Duration::from_secs(60);
+const STALE_SESSION_AGE: Duration = Duration::from_secs(10 * 60);
+
+#[derive(Debug)]
+struct CachedPack {
+    dir: PathBuf,
+    pack_path: PathBuf,
+    idx_path: PathBuf,
+    pack_bytes: u64,
+    total_bytes: u64,
+    /// Keeps an over-capacity staging directory alive until it is linked into
+    /// the request workspace. Normal cached entries are atomically renamed and
+    /// do not need a temporary owner.
+    _temporary: Option<tempfile::TempDir>,
+}
+
+impl CachedPack {
+    fn from_dir(dir: PathBuf, digest: &str) -> Option<Self> {
+        let pack_path = dir.join(format!("pack-{digest}.pack"));
+        let idx_path = dir.join(format!("pack-{digest}.idx"));
+        let pack_bytes = regular_file_len(&pack_path)?;
+        let idx_bytes = regular_file_len(&idx_path)?;
+        Some(Self {
+            dir,
+            pack_path,
+            idx_path,
+            pack_bytes,
+            total_bytes: pack_bytes.saturating_add(idx_bytes),
+            _temporary: None,
+        })
+    }
+}
+
+struct CacheRecord {
+    entry: Arc<CachedPack>,
+    last_used: u64,
+}
+
+#[derive(Default)]
+struct CacheState {
+    entries: HashMap<String, CacheRecord>,
+    total_bytes: u64,
+    tick: u64,
+}
+
+/// Process-local, byte-bounded cache of immutable pack/index pairs.
+pub struct GitPackCache {
+    _session_dir: tempfile::TempDir,
+    heartbeat_task: Option<tokio::task::JoinHandle<()>>,
+    root: PathBuf,
+    max_bytes: u64,
+    population_semaphore: tokio::sync::Semaphore,
+    state: Mutex<CacheState>,
+    flights: DashMap<String, Arc<PopulationFlight>>,
+}
+
+struct PopulationFlight {
+    lock: tokio::sync::Mutex<()>,
+    participants: AtomicUsize,
+}
+
+struct FlightParticipant<'a> {
+    cache: &'a GitPackCache,
+    digest: &'a str,
+    flight: Arc<PopulationFlight>,
+}
+
+struct PopulationPermit<'a> {
+    _permit: tokio::sync::SemaphorePermit<'a>,
+}
+
+impl Drop for PopulationPermit<'_> {
+    fn drop(&mut self) {
+        metrics::gauge!("buzz_git_pack_cache_populations_active").decrement(1.0);
+    }
+}
+
+impl Drop for FlightParticipant<'_> {
+    fn drop(&mut self) {
+        if self.flight.participants.fetch_sub(1, Ordering::AcqRel) == 1 {
+            self.cache.remove_flight(self.digest, &self.flight);
+        }
+    }
+}
+
+impl GitPackCache {
+    /// Create an isolated process-lifetime cache beneath `cache_parent`.
+    pub fn new(
+        cache_parent: &Path,
+        max_bytes: u64,
+        max_concurrent_populations: usize,
+    ) -> Result<Self, String> {
+        if max_concurrent_populations == 0 {
+            return Err("git pack cache population concurrency must be positive".to_string());
+        }
+        std::fs::create_dir_all(cache_parent)
+            .map_err(|error| format!("create git pack cache {cache_parent:?}: {error}"))?;
+        if std::fs::symlink_metadata(cache_parent)
+            .map_err(|error| format!("stat git pack cache {cache_parent:?}: {error}"))?
+            .file_type()
+            .is_symlink()
+        {
+            return Err(format!(
+                "git pack cache path must not be a symlink: {cache_parent:?}"
+            ));
+        }
+        cleanup_stale_sessions(cache_parent);
+        let session_dir = Builder::new()
+            .prefix("session-")
+            .tempdir_in(cache_parent)
+            .map_err(|error| format!("create git pack cache session: {error}"))?;
+        let root = session_dir.path().to_path_buf();
+        let heartbeat_path = root.join(".heartbeat");
+        std::fs::write(&heartbeat_path, b"")
+            .map_err(|error| format!("create git pack cache heartbeat: {error}"))?;
+        let heartbeat_task = tokio::runtime::Handle::try_current().ok().map(|runtime| {
+            runtime.spawn(async move {
+                let mut interval = tokio::time::interval(HEARTBEAT_INTERVAL);
+                interval.tick().await;
+                loop {
+                    interval.tick().await;
+                    if tokio::fs::write(&heartbeat_path, b"").await.is_err() {
+                        break;
+                    }
+                }
+            })
+        });
+        let cache = Self {
+            _session_dir: session_dir,
+            heartbeat_task,
+            root,
+            max_bytes,
+            population_semaphore: tokio::sync::Semaphore::new(max_concurrent_populations),
+            state: Mutex::new(CacheState::default()),
+            flights: DashMap::new(),
+        };
+        emit_size_metrics(&CacheState::default());
+        Ok(cache)
+    }
+
+    /// Materialize one verified pack/index pair into a request workspace.
+    ///
+    /// Concurrent misses for the same digest share one population flight.
+    pub async fn materialize_pack(
+        &self,
+        store: &GitStore,
+        object_key: &str,
+        digest: &str,
+        destination: &Path,
+        max_pack_bytes: u64,
+    ) -> Result<u64, HydrateError> {
+        validate_digest(digest)?;
+        if let Some(entry) = self.lookup(digest) {
+            metrics::counter!("buzz_git_pack_cache_lookups_total", "result" => "hit").increment(1);
+            let pack_bytes = entry.pack_bytes;
+            if let Ok(()) = install_entry(&entry, digest, destination).await {
+                drop(entry);
+                self.prune();
+                return Ok(pack_bytes);
+            }
+            self.invalidate(digest);
+        }
+
+        let (flight, joined) = match self.flights.entry(digest.to_string()) {
+            Entry::Occupied(entry) => {
+                entry.get().participants.fetch_add(1, Ordering::Relaxed);
+                (Arc::clone(entry.get()), true)
+            }
+            Entry::Vacant(entry) => {
+                let flight = Arc::new(PopulationFlight {
+                    lock: tokio::sync::Mutex::new(()),
+                    participants: AtomicUsize::new(1),
+                });
+                entry.insert(Arc::clone(&flight));
+                (flight, false)
+            }
+        };
+        metrics::counter!(
+            "buzz_git_pack_cache_lookups_total",
+            "result" => if joined { "coalesced" } else { "miss" }
+        )
+        .increment(1);
+
+        let flight_participant = FlightParticipant {
+            cache: self,
+            digest,
+            flight: Arc::clone(&flight),
+        };
+        let guard = flight.lock.lock().await;
+        let result = if let Some(entry) = self.lookup(digest) {
+            install_entry(&entry, digest, destination)
+                .await
+                .map(|()| entry.pack_bytes)
+        } else {
+            let started_at = Instant::now();
+            let populated = self
+                .populate(store, object_key, digest, max_pack_bytes)
+                .await;
+            let outcome = match &populated {
+                Ok(entry) if entry._temporary.is_some() => "bypass",
+                Ok(_) => "success",
+                Err(_) => "error",
+            };
+            metrics::histogram!(
+                "buzz_git_pack_cache_populate_seconds",
+                "outcome" => outcome
+            )
+            .record(started_at.elapsed().as_secs_f64());
+            match populated {
+                Ok(entry) => install_entry(&entry, digest, destination)
+                    .await
+                    .map(|()| entry.pack_bytes),
+                Err(error) => Err(error),
+            }
+        };
+        drop(guard);
+        drop(flight_participant);
+        self.prune();
+        result
+    }
+
+    fn lookup(&self, digest: &str) -> Option<Arc<CachedPack>> {
+        let mut state = self.state.lock().unwrap_or_else(|error| error.into_inner());
+        let next_tick = state.tick.saturating_add(1);
+        state.tick = next_tick;
+        let record = state.entries.get_mut(digest)?;
+        if !record.entry.pack_path.is_file() || !record.entry.idx_path.is_file() {
+            return None;
+        }
+        record.last_used = next_tick;
+        Some(Arc::clone(&record.entry))
+    }
+
+    async fn populate(
+        &self,
+        store: &GitStore,
+        object_key: &str,
+        digest: &str,
+        max_pack_bytes: u64,
+    ) -> Result<Arc<CachedPack>, HydrateError> {
+        let wait_started_at = Instant::now();
+        let permit = self
+            .population_semaphore
+            .acquire()
+            .await
+            .map_err(|_| HydrateError::Hydrate("pack cache population closed".to_string()))?;
+        metrics::histogram!("buzz_git_pack_cache_population_wait_seconds")
+            .record(wait_started_at.elapsed().as_secs_f64());
+        metrics::gauge!("buzz_git_pack_cache_populations_active").increment(1.0);
+        let _population_permit = PopulationPermit { _permit: permit };
+        let shard = self.root.join(&digest[..2]);
+        tokio::fs::create_dir_all(&shard)
+            .await
+            .map_err(|error| HydrateError::Hydrate(format!("create pack cache shard: {error}")))?;
+        let staging = Builder::new()
+            .prefix(".staging-")
+            .tempdir_in(&shard)
+            .map_err(|error| {
+                HydrateError::Hydrate(format!("create pack cache staging directory: {error}"))
+            })?;
+        let pack_path = staging.path().join(format!("pack-{digest}.pack"));
+        let bytes = get_verified_limited(store, object_key, digest, max_pack_bytes).await?;
+        let pack_bytes = u64::try_from(bytes.len()).unwrap_or(u64::MAX);
+        tokio::fs::write(&pack_path, &bytes)
+            .await
+            .map_err(|error| HydrateError::Hydrate(format!("write cached pack: {error}")))?;
+        install_or_generate_idx(store, staging.path(), digest, &pack_path, max_pack_bytes).await?;
+        let idx_path = pack_path.with_extension("idx");
+        let idx_bytes = tokio::fs::metadata(&idx_path)
+            .await
+            .map_err(|error| HydrateError::Hydrate(format!("stat cached idx: {error}")))?
+            .len();
+        let total_bytes = pack_bytes.saturating_add(idx_bytes);
+
+        if self.max_bytes == 0 || total_bytes > self.max_bytes {
+            metrics::counter!("buzz_git_pack_cache_bypasses_total").increment(1);
+            return Ok(Arc::new(CachedPack {
+                dir: staging.path().to_path_buf(),
+                pack_path,
+                idx_path,
+                pack_bytes,
+                total_bytes,
+                _temporary: Some(staging),
+            }));
+        }
+
+        let final_dir = shard.join(digest);
+        if let Some(entry) = CachedPack::from_dir(final_dir.clone(), digest) {
+            let entry = Arc::new(entry);
+            self.insert(digest.to_string(), Arc::clone(&entry));
+            return Ok(entry);
+        }
+        if final_dir.exists() {
+            let _ = tokio::fs::remove_dir_all(&final_dir).await;
+        }
+        if let Err(error) = tokio::fs::rename(staging.path(), &final_dir).await {
+            // Another relay process sharing the scratch volume may have won
+            // the same content-addressed publication race.
+            if let Some(entry) = CachedPack::from_dir(final_dir.clone(), digest) {
+                let entry = Arc::new(entry);
+                self.insert(digest.to_string(), Arc::clone(&entry));
+                return Ok(entry);
+            }
+            return Err(HydrateError::Hydrate(format!(
+                "publish cached pack: {error}"
+            )));
+        }
+        let entry = Arc::new(CachedPack::from_dir(final_dir, digest).ok_or_else(|| {
+            HydrateError::Hydrate("published pack cache entry is incomplete".to_string())
+        })?);
+        self.insert(digest.to_string(), Arc::clone(&entry));
+        Ok(entry)
+    }
+
+    fn insert(&self, digest: String, entry: Arc<CachedPack>) {
+        let mut state = self.state.lock().unwrap_or_else(|error| error.into_inner());
+        state.tick = state.tick.saturating_add(1);
+        let last_used = state.tick;
+        if let Some(previous) = state.entries.remove(&digest) {
+            state.total_bytes = state.total_bytes.saturating_sub(previous.entry.total_bytes);
+        }
+        state.total_bytes = state.total_bytes.saturating_add(entry.total_bytes);
+        state
+            .entries
+            .insert(digest, CacheRecord { entry, last_used });
+        emit_size_metrics(&state);
+    }
+
+    fn invalidate(&self, digest: &str) {
+        let removed = {
+            let mut state = self.state.lock().unwrap_or_else(|error| error.into_inner());
+            let removed = state.entries.remove(digest);
+            if let Some(record) = &removed {
+                state.total_bytes = state.total_bytes.saturating_sub(record.entry.total_bytes);
+            }
+            emit_size_metrics(&state);
+            removed
+        };
+        if let Some(record) = removed {
+            if Arc::strong_count(&record.entry) == 1 {
+                let _ = std::fs::remove_dir_all(&record.entry.dir);
+            }
+        }
+    }
+
+    fn prune(&self) {
+        let mut removed = Vec::new();
+        {
+            let mut state = self.state.lock().unwrap_or_else(|error| error.into_inner());
+            while state.total_bytes > self.max_bytes {
+                let candidate = state
+                    .entries
+                    .iter()
+                    .filter(|(_, record)| Arc::strong_count(&record.entry) == 1)
+                    .min_by_key(|(_, record)| record.last_used)
+                    .map(|(digest, _)| digest.clone());
+                let Some(digest) = candidate else {
+                    break;
+                };
+                if let Some(record) = state.entries.remove(&digest) {
+                    state.total_bytes = state.total_bytes.saturating_sub(record.entry.total_bytes);
+                    removed.push(record.entry.dir.clone());
+                    metrics::counter!("buzz_git_pack_cache_evictions_total").increment(1);
+                }
+            }
+            emit_size_metrics(&state);
+        }
+        for dir in removed {
+            let _ = std::fs::remove_dir_all(dir);
+        }
+    }
+
+    fn remove_flight(&self, digest: &str, flight: &Arc<PopulationFlight>) {
+        if let Entry::Occupied(entry) = self.flights.entry(digest.to_string()) {
+            if Arc::ptr_eq(entry.get(), flight) && flight.participants.load(Ordering::Acquire) == 0
+            {
+                entry.remove();
+            }
+        }
+    }
+
+    #[cfg(test)]
+    fn flight(&self, digest: &str) -> (Arc<PopulationFlight>, bool) {
+        match self.flights.entry(digest.to_string()) {
+            Entry::Occupied(entry) => {
+                entry.get().participants.fetch_add(1, Ordering::Relaxed);
+                (Arc::clone(entry.get()), true)
+            }
+            Entry::Vacant(entry) => {
+                let flight = Arc::new(PopulationFlight {
+                    lock: tokio::sync::Mutex::new(()),
+                    participants: AtomicUsize::new(1),
+                });
+                entry.insert(Arc::clone(&flight));
+                (flight, false)
+            }
+        }
+    }
+}
+
+impl Drop for GitPackCache {
+    fn drop(&mut self) {
+        if let Some(task) = self.heartbeat_task.take() {
+            task.abort();
+        }
+    }
+}
+
+async fn install_entry(
+    entry: &CachedPack,
+    digest: &str,
+    destination: &Path,
+) -> Result<(), HydrateError> {
+    let destination_pack = destination.join(format!("pack-{digest}.pack"));
+    let destination_idx = destination.join(format!("pack-{digest}.idx"));
+    if tokio::fs::hard_link(&entry.pack_path, &destination_pack)
+        .await
+        .is_ok()
+        && tokio::fs::hard_link(&entry.idx_path, &destination_idx)
+            .await
+            .is_ok()
+    {
+        return Ok(());
+    }
+
+    let _ = tokio::fs::remove_file(&destination_pack).await;
+    let _ = tokio::fs::remove_file(&destination_idx).await;
+    tokio::fs::copy(&entry.pack_path, &destination_pack)
+        .await
+        .map_err(|error| HydrateError::Hydrate(format!("copy cached pack: {error}")))?;
+    if let Err(error) = tokio::fs::copy(&entry.idx_path, &destination_idx).await {
+        let _ = tokio::fs::remove_file(&destination_pack).await;
+        return Err(HydrateError::Hydrate(format!("copy cached idx: {error}")));
+    }
+    metrics::counter!("buzz_git_pack_cache_copy_fallbacks_total").increment(1);
+    Ok(())
+}
+
+fn validate_digest(digest: &str) -> Result<(), HydrateError> {
+    if digest.len() == 64
+        && digest
+            .chars()
+            .all(|character| character.is_ascii_hexdigit())
+    {
+        Ok(())
+    } else {
+        Err(HydrateError::Hydrate(
+            "pack cache digest is malformed".to_string(),
+        ))
+    }
+}
+
+fn regular_file_len(path: &Path) -> Option<u64> {
+    let metadata = std::fs::symlink_metadata(path).ok()?;
+    metadata.file_type().is_file().then_some(metadata.len())
+}
+
+fn emit_size_metrics(state: &CacheState) {
+    metrics::gauge!("buzz_git_pack_cache_bytes").set(state.total_bytes as f64);
+    metrics::gauge!("buzz_git_pack_cache_entries").set(state.entries.len() as f64);
+}
+
+fn cleanup_stale_sessions(cache_parent: &Path) {
+    cleanup_sessions_older_than(cache_parent, STALE_SESSION_AGE);
+}
+
+fn cleanup_sessions_older_than(cache_parent: &Path, max_age: Duration) {
+    let Ok(entries) = std::fs::read_dir(cache_parent) else {
+        return;
+    };
+    for entry in entries.flatten() {
+        let name = entry.file_name();
+        if !name.to_string_lossy().starts_with("session-")
+            || !entry.file_type().is_ok_and(|file_type| file_type.is_dir())
+        {
+            continue;
+        }
+        let heartbeat = entry.path().join(".heartbeat");
+        let modified = std::fs::symlink_metadata(&heartbeat)
+            .or_else(|_| entry.metadata())
+            .and_then(|metadata| metadata.modified());
+        let stale = modified
+            .ok()
+            .and_then(|modified| SystemTime::now().duration_since(modified).ok())
+            .is_some_and(|age| age > max_age);
+        if stale {
+            let _ = std::fs::remove_dir_all(entry.path());
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn digest(character: char) -> String {
+        std::iter::repeat_n(character, 64).collect()
+    }
+
+    fn write_entry(root: &Path, digest: &str, pack_bytes: usize, idx_bytes: usize) -> PathBuf {
+        let dir = root.join(&digest[..2]).join(digest);
+        std::fs::create_dir_all(&dir).expect("entry dir");
+        std::fs::write(dir.join(format!("pack-{digest}.pack")), vec![1; pack_bytes]).expect("pack");
+        std::fs::write(dir.join(format!("pack-{digest}.idx")), vec![2; idx_bytes]).expect("idx");
+        dir
+    }
+
+    #[test]
+    fn entries_are_bounded_by_total_bytes() {
+        let scratch = tempfile::TempDir::new().expect("scratch");
+        let cache = GitPackCache::new(scratch.path(), 8, 2).expect("cache");
+        let first = digest('a');
+        let second = digest('b');
+        let first_dir = write_entry(&cache.root, &first, 6, 2);
+        let second_dir = write_entry(&cache.root, &second, 6, 2);
+        cache.insert(
+            first.clone(),
+            Arc::new(CachedPack::from_dir(first_dir, &first).expect("first entry")),
+        );
+        cache.insert(
+            second.clone(),
+            Arc::new(CachedPack::from_dir(second_dir, &second).expect("second entry")),
+        );
+        cache.prune();
+
+        let state = cache.state.lock().expect("state");
+        assert_eq!(state.entries.len(), 1);
+        assert_eq!(state.total_bytes, 8);
+    }
+
+    #[test]
+    fn cache_sessions_are_process_isolated() {
+        let parent = tempfile::TempDir::new().expect("parent");
+        let first = GitPackCache::new(parent.path(), 8, 2).expect("first");
+        let second = GitPackCache::new(parent.path(), 8, 2).expect("second");
+
+        assert_ne!(first.root, second.root);
+        assert!(first.root.starts_with(parent.path()));
+        assert!(second.root.starts_with(parent.path()));
+    }
+
+    #[test]
+    fn abandoned_sessions_are_removed_after_grace_period() {
+        let parent = tempfile::TempDir::new().expect("parent");
+        let abandoned = parent.path().join("session-abandoned");
+        std::fs::create_dir(&abandoned).expect("abandoned");
+        std::fs::write(abandoned.join(".heartbeat"), b"").expect("heartbeat");
+        std::thread::sleep(Duration::from_millis(2));
+
+        cleanup_sessions_older_than(parent.path(), Duration::ZERO);
+
+        assert!(!abandoned.exists());
+    }
+
+    #[cfg(unix)]
+    #[test]
+    fn cache_parent_must_not_be_a_symlink() {
+        let parent = tempfile::TempDir::new().expect("parent");
+        let target = parent.path().join("target");
+        let link = parent.path().join("link");
+        std::fs::create_dir(&target).expect("target");
+        std::os::unix::fs::symlink(&target, &link).expect("link");
+
+        assert!(GitPackCache::new(&link, 8, 2).is_err());
+    }
+
+    #[test]
+    fn cache_digest_cannot_escape_cache_root() {
+        for digest in ["../pack", "/absolute", "g", &"a".repeat(63)] {
+            assert!(validate_digest(digest).is_err(), "{digest:?}");
+        }
+        assert!(validate_digest(&"a".repeat(64)).is_ok());
+    }
+
+    #[tokio::test]
+    async fn concurrent_digest_requests_share_one_flight() {
+        let scratch = tempfile::TempDir::new().expect("scratch");
+        let cache = GitPackCache::new(scratch.path(), 1024, 2).expect("cache");
+        let digest = digest('c');
+        let (leader, joined) = cache.flight(&digest);
+        assert!(!joined);
+        let leader_participant = FlightParticipant {
+            cache: &cache,
+            digest: &digest,
+            flight: Arc::clone(&leader),
+        };
+        let leader_guard = leader.lock.lock().await;
+
+        let (waiter, joined) = cache.flight(&digest);
+        assert!(joined);
+        assert!(Arc::ptr_eq(&leader, &waiter));
+        let waiter_participant = FlightParticipant {
+            cache: &cache,
+            digest: &digest,
+            flight: Arc::clone(&waiter),
+        };
+        assert!(
+            tokio::time::timeout(std::time::Duration::from_millis(10), waiter.lock.lock())
+                .await
+                .is_err()
+        );
+
+        drop(leader_guard);
+        assert!(
+            tokio::time::timeout(std::time::Duration::from_secs(1), waiter.lock.lock())
+                .await
+                .is_ok()
+        );
+        drop(waiter_participant);
+        assert!(cache.flights.contains_key(&digest));
+        drop(leader_participant);
+        assert!(!cache.flights.contains_key(&digest));
+    }
+
+    #[tokio::test]
+    async fn cancelled_waiter_keeps_active_flight_registered() {
+        let scratch = tempfile::TempDir::new().expect("scratch");
+        let cache = GitPackCache::new(scratch.path(), 1024, 2).expect("cache");
+        let digest = digest('e');
+        let (leader, _) = cache.flight(&digest);
+        let leader_participant = FlightParticipant {
+            cache: &cache,
+            digest: &digest,
+            flight: Arc::clone(&leader),
+        };
+        let leader_guard = leader.lock.lock().await;
+        let (waiter, joined) = cache.flight(&digest);
+        assert!(joined);
+        let waiter_participant = FlightParticipant {
+            cache: &cache,
+            digest: &digest,
+            flight: waiter,
+        };
+
+        drop(waiter_participant);
+        assert!(cache.flights.contains_key(&digest));
+
+        drop(leader_guard);
+        drop(leader_participant);
+        assert!(!cache.flights.contains_key(&digest));
+    }
+
+    #[tokio::test]
+    async fn cached_pair_is_linked_into_request_workspace() {
+        let scratch = tempfile::TempDir::new().expect("scratch");
+        let cache = GitPackCache::new(scratch.path(), 1024, 2).expect("cache");
+        let digest = digest('d');
+        let source = write_entry(&cache.root, &digest, 3, 2);
+        cache.insert(
+            digest.clone(),
+            Arc::new(CachedPack::from_dir(source.clone(), &digest).expect("entry")),
+        );
+        let entry = cache.lookup(&digest).expect("entry");
+        let destination = tempfile::TempDir::new().expect("destination");
+
+        install_entry(&entry, &digest, destination.path())
+            .await
+            .expect("install");
+
+        assert_eq!(
+            std::fs::read(destination.path().join(format!("pack-{digest}.pack")))
+                .expect("installed pack"),
+            vec![1; 3]
+        );
+        assert!(source.is_dir());
+    }
+}

crates/buzz-relay/src/api/git/store.rs

@@ -320,9 +320,13 @@ impl GitStore {
     /// A missing idx is a cache miss, not a hydrate failure; callers should
     /// regenerate with `git index-pack`. Other backend failures are surfaced so
     /// callers can decide whether to fall back or fail.
-    pub async fn get_idx(&self, pack_digest: &str) -> Result<Option<Bytes>, StoreError> {
+    pub async fn get_idx(
+        &self,
+        pack_digest: &str,
+        max_bytes: u64,
+    ) -> Result<Option<Bytes>, StoreError> {
         let key = Self::idx_key_for_pack_digest(pack_digest)?;
-        match self.get(&key).await {
+        match self.get_limited(&key, max_bytes).await {
             Ok(bytes) => Ok(Some(bytes)),
             Err(StoreError::NotFound(_)) => Ok(None),
             Err(e) => Err(e),
@@ -384,6 +388,22 @@ impl GitStore {
         expected_digest: &str,
         max_bytes: u64,
     ) -> Result<Bytes, StoreError> {
+        let bytes = self.get_limited(key, max_bytes).await?;
+        let mut hasher = Sha256::new();
+        hasher.update(&bytes);
+        let actual = hex::encode(hasher.finalize());
+        if actual != expected_digest {
+            return Err(StoreError::DigestMismatch {
+                key: key.into(),
+                expected: expected_digest.into(),
+                actual,
+            });
+        }
+        Ok(bytes)
+    }
+
+    /// GET an object after rejecting bodies larger than `max_bytes`.
+    pub async fn get_limited(&self, key: &str, max_bytes: u64) -> Result<Bytes, StoreError> {
         let (head, status) = self.bucket.head_object(key).await.map_err(|e| match e {
             S3Error::HttpFailWithBody(404, _) => StoreError::NotFound(key.into()),
             other => StoreError::Backend(other),
@@ -408,7 +428,7 @@ impl GitStore {
             }
         }
 
-        let bytes = self.get_verified(key, expected_digest).await?;
+        let bytes = self.get(key).await?;
         let size = u64::try_from(bytes.len()).unwrap_or(u64::MAX);
         if size > max_bytes {
             return Err(StoreError::ObjectTooLarge {

crates/buzz-relay/src/api/git/transport.rs

@@ -32,6 +32,7 @@ use super::cas_publish::{cas_publish, CasError, ParentState, PublishLimits};
 use super::hook::install_hook;
 use super::hydrate::{
     hydrate_for_read, hydrate_for_write, load_manifest_for_read, HydrateError, HydratedRepo,
+    HydrationOptions,
 };
 use super::manifest_event::{build_ref_state_event, RefStateInputs};
 use crate::state::AppState;
@@ -598,9 +599,12 @@ async fn info_refs_subprocess(
         tenant,
         &params.owner,
         &params.repo,
-        &state.config.git_repo_path,
+        HydrationOptions {
-        state.config.git_max_pack_bytes,
+            pack_cache: &state.git_pack_cache,
-        state.config.git_max_repo_bytes,
+            scratch_dir: &state.config.git_repo_path,
+            max_pack_bytes: state.config.git_max_pack_bytes,
+            max_repo_bytes: state.config.git_max_repo_bytes,
+        },
     )
     .await
     {
@@ -730,9 +734,12 @@ pub async fn upload_pack(
         &auth.tenant,
         &params.owner,
         &params.repo,
-        &state.config.git_repo_path,
+        HydrationOptions {
-        state.config.git_max_pack_bytes,
+            pack_cache: &state.git_pack_cache,
-        state.config.git_max_repo_bytes,
+            scratch_dir: &state.config.git_repo_path,
+            max_pack_bytes: state.config.git_max_pack_bytes,
+            max_repo_bytes: state.config.git_max_repo_bytes,
+        },
     )
     .await
     {
@@ -816,9 +823,12 @@ pub async fn receive_pack(
         &auth.tenant,
         &params.owner,
         &params.repo,
-        &state.config.git_repo_path,
+        HydrationOptions {
-        state.config.git_max_pack_bytes,
+            pack_cache: &state.git_pack_cache,
-        state.config.git_max_repo_bytes,
+            scratch_dir: &state.config.git_repo_path,
+            max_pack_bytes: state.config.git_max_pack_bytes,
+            max_repo_bytes: state.config.git_max_repo_bytes,
+        },
     )
     .await
     .map_err(|e| hydrate_error_to_response(&params.owner, &params.repo, e))?;

crates/buzz-relay/src/config.rs

@@ -202,13 +202,16 @@ pub struct Config {
     /// 60 seconds after the last message.
     pub ephemeral_ttl_override: Option<i32>,
 
-    /// Root directory for the relay's local git scratch. No per-repo bare repos
+    /// Root directory for the relay's local git scratch. No authoritative
-    /// or persistent git state live here — runtime reads/writes hydrate
+    /// repository state lives here — runtime reads/writes hydrate ephemeral
-    /// ephemeral repos from object storage per request, and all temporary Git
+    /// repos from object storage per request. Temporary workspaces, buffered
-    /// workspaces and buffered subprocess output are created beneath this path.
+    /// subprocess output, and the disposable immutable pack cache live below
+    /// this path.
     /// Repo-name uniqueness lives in Postgres (`git_repo_names`), not on disk,
     /// so this directory need not be persistent or shared across replicas.
     pub git_repo_path: std::path::PathBuf,
+    /// Parent directory for process-isolated immutable pack cache sessions.
+    pub git_pack_cache_path: std::path::PathBuf,
     /// Maximum pack file size for git push (bytes). Default: 500 MB.
     pub git_max_pack_bytes: u64,
     /// Maximum total bytes materialized for one git repo request. Default: 1 GB.
@@ -216,6 +219,11 @@ pub struct Config {
     /// This bounds clone/fetch hydration work across a repo's historical pack
     /// set rather than only bounding one incoming push body.
     pub git_max_repo_bytes: u64,
+    /// Maximum bytes retained in the process-local immutable pack/index cache.
+    /// Zero disables retention while preserving request-local hydration.
+    pub git_pack_cache_max_bytes: u64,
+    /// Maximum pack digests populated concurrently in one relay process.
+    pub git_pack_cache_max_concurrent_populations: usize,
     /// Maximum number of repos per pubkey. Default: 100.
     pub git_max_repos_per_pubkey: u32,
     /// Maximum concurrent git subprocess operations. Default: 20.
@@ -368,11 +376,18 @@ fn parse_optional_bool(name: &str) -> Result<bool, ConfigError> {
 
 fn ensure_git_repo_path(
     raw: impl Into<std::path::PathBuf>,
+) -> Result<std::path::PathBuf, ConfigError> {
+    ensure_git_path("BUZZ_GIT_REPO_PATH", raw)
+}
+
+fn ensure_git_path(
+    setting: &str,
+    raw: impl Into<std::path::PathBuf>,
 ) -> Result<std::path::PathBuf, ConfigError> {
     let git_repo_path = raw.into();
     if let Err(e) = std::fs::create_dir_all(&git_repo_path) {
         return Err(ConfigError::InvalidValue(format!(
-            "BUZZ_GIT_REPO_PATH={} could not be created: {e}",
+            "{setting}={} could not be created: {e}",
             git_repo_path.display()
         )));
     }
@@ -677,6 +692,12 @@ impl Config {
         let git_repo_path = ensure_git_repo_path(
             std::env::var("BUZZ_GIT_REPO_PATH").unwrap_or_else(|_| "./repos".to_string()),
         )?;
+        let git_pack_cache_path = ensure_git_path(
+            "BUZZ_GIT_PACK_CACHE_PATH",
+            std::env::var("BUZZ_GIT_PACK_CACHE_PATH")
+                .map(std::path::PathBuf::from)
+                .unwrap_or_else(|_| git_repo_path.join(".pack-cache")),
+        )?;
         let git_max_pack_bytes: u64 = std::env::var("BUZZ_GIT_MAX_PACK_BYTES")
             .ok()
             .and_then(|v| v.parse().ok())
@@ -685,6 +706,16 @@ impl Config {
             .ok()
             .and_then(|v| v.parse().ok())
             .unwrap_or_else(|| git_max_pack_bytes.saturating_mul(2)); // 1 GB at defaults
+        let git_pack_cache_max_bytes: u64 = std::env::var("BUZZ_GIT_PACK_CACHE_MAX_BYTES")
+            .ok()
+            .and_then(|v| v.parse().ok())
+            .unwrap_or_else(|| git_max_repo_bytes.saturating_mul(5)); // 5 GB at defaults
+        let git_pack_cache_max_concurrent_populations: usize =
+            std::env::var("BUZZ_GIT_PACK_CACHE_MAX_CONCURRENT_POPULATIONS")
+                .ok()
+                .and_then(|v| v.parse().ok())
+                .filter(|value| *value > 0)
+                .unwrap_or(2);
         let git_max_repos_per_pubkey: u32 = std::env::var("BUZZ_GIT_MAX_REPOS_PER_PUBKEY")
             .ok()
             .and_then(|v| v.parse().ok())
@@ -862,8 +893,11 @@ impl Config {
             audit_enabled,
             ephemeral_ttl_override,
             git_repo_path,
+            git_pack_cache_path,
             git_max_pack_bytes,
             git_max_repo_bytes,
+            git_pack_cache_max_bytes,
+            git_pack_cache_max_concurrent_populations,
             git_max_repos_per_pubkey,
             git_max_concurrent_ops,
             git_hook_hmac_secret,

crates/buzz-relay/src/metrics.rs

@@ -87,6 +87,16 @@ pub fn install(port: u16, gauge_idle_timeout_secs: u64) {
             &GIT_DURATION_BUCKETS_S,
         )
         .expect("valid git stream duration bucket boundaries")
+        .set_buckets_for_metric(
+            Matcher::Full("buzz_git_pack_cache_populate_seconds".to_owned()),
+            &GIT_DURATION_BUCKETS_S,
+        )
+        .expect("valid git cache population duration bucket boundaries")
+        .set_buckets_for_metric(
+            Matcher::Full("buzz_git_pack_cache_population_wait_seconds".to_owned()),
+            &GIT_DURATION_BUCKETS_S,
+        )
+        .expect("valid git cache population wait bucket boundaries")
         .set_buckets_for_metric(
             Matcher::Full("buzz_git_hydrate_bytes".to_owned()),
             &GIT_BYTES_BUCKETS,

crates/buzz-relay/src/state.rs

@@ -505,6 +505,10 @@ pub struct AppState {
     /// CAS-guarded manifest pointer). This is the durable git source of truth;
     /// see `api::git::store` and `docs/git-on-object-storage.md`.
     pub git_store: crate::api::git::store::GitStore,
+    /// Process-local, byte-bounded cache of immutable Git pack/index pairs.
+    /// Object storage remains authoritative; this only avoids repeated reads
+    /// and index generation for content-addressed packs.
+    pub git_pack_cache: Arc<crate::api::git::pack_cache::GitPackCache>,
     /// Audio relay room manager — tracks active huddle audio rooms.
     pub audio_rooms: Arc<AudioRoomManager>,
     /// Set to `true` on SIGTERM — readiness probe returns 503.
@@ -631,6 +635,14 @@ impl AppState {
             &config.media.s3_region,
         )
         .expect("media storage was already constructed with this S3 config");
+        let git_pack_cache = Arc::new(
+            crate::api::git::pack_cache::GitPackCache::new(
+                &config.git_pack_cache_path,
+                config.git_pack_cache_max_bytes,
+                config.git_pack_cache_max_concurrent_populations,
+            )
+            .expect("git pack cache path must be available"),
+        );
         let nip98_replay: Arc<dyn Nip98ReplayGuard> =
             Arc::new(RedisNip98ReplayGuard::new(redis_pool.clone()));
         let admission_rate_limiter = Arc::new(RedisRateLimiter::new(redis_pool.clone()));
@@ -685,6 +697,7 @@ impl AppState {
             audit_tx: audit_enabled.then_some(audit_tx),
             media_storage: Arc::new(media_storage),
             git_store,
+            git_pack_cache,
             audio_rooms: Arc::new(AudioRoomManager::new()),
             shutting_down: Arc::new(AtomicBool::new(false)),
             started_at: Instant::now(),

deploy/charts/buzz/templates/deployment.yaml

@@ -145,6 +145,9 @@ spec:
             # ── Git ──────────────────────────────────────────────────
             - { name: BUZZ_GIT_REPO_PATH,         value: {{ .Values.persistence.git.mountPath | quote }} }
             - { name: BUZZ_GIT_MAX_PACK_BYTES,    value: {{ .Values.git.maxPackBytes | quote }} }
+            - { name: BUZZ_GIT_PACK_CACHE_PATH,   value: {{ .Values.git.packCachePath | quote }} }
+            - { name: BUZZ_GIT_PACK_CACHE_MAX_BYTES, value: {{ .Values.git.packCacheMaxBytes | quote }} }
+            - { name: BUZZ_GIT_PACK_CACHE_MAX_CONCURRENT_POPULATIONS, value: {{ .Values.git.packCacheMaxConcurrentPopulations | quote }} }
             - { name: BUZZ_GIT_MAX_REPOS_PER_PUBKEY, value: {{ .Values.git.maxReposPerPubkey | quote }} }
             - { name: BUZZ_GIT_MAX_CONCURRENT_OPS, value: {{ .Values.git.maxConcurrentOps | quote }} }
 
@@ -221,6 +224,7 @@ spec:
 
           volumeMounts:
             - { name: git-repos, mountPath: {{ .Values.persistence.git.mountPath | quote }} }
+            - { name: git-pack-cache, mountPath: {{ .Values.git.packCachePath | quote }} }
 
       volumes:
         - name: git-repos
@@ -231,3 +235,6 @@ spec:
           emptyDir:
             sizeLimit: {{ .Values.persistence.git.size | quote }}
         {{- end }}
+        - name: git-pack-cache
+          emptyDir:
+            sizeLimit: {{ .Values.git.packCacheVolumeSize | quote }}

deploy/charts/buzz/values.schema.json

@@ -199,6 +199,10 @@
       "additionalProperties": false,
       "properties": {
         "maxPackBytes": { "type": "integer", "minimum": 1 },
+        "packCachePath": { "type": "string", "minLength": 1 },
+        "packCacheMaxBytes": { "type": "integer", "minimum": 0 },
+        "packCacheMaxConcurrentPopulations": { "type": "integer", "minimum": 1 },
+        "packCacheVolumeSize": { "type": "string", "pattern": "^[0-9]+(\\.[0-9]+)?(E|P|T|G|M|K|Ei|Pi|Ti|Gi|Mi|Ki)?$" },
         "maxReposPerPubkey": { "type": "integer", "minimum": 1 },
         "maxConcurrentOps": { "type": "integer", "minimum": 1 }
       }

deploy/charts/buzz/values.yaml

@@ -323,6 +323,10 @@ minio:
 # ── Git server config ────────────────────────────────────────────────────────
 git:
   maxPackBytes: 524288000        # 500 MiB
+  packCachePath: /var/cache/buzz/git-packs
+  packCacheMaxBytes: 5368709120  # 5 GiB
+  packCacheMaxConcurrentPopulations: 2
+  packCacheVolumeSize: 7Gi       # per-pod emptyDir; includes cold-population staging
   maxReposPerPubkey: 100
   maxConcurrentOps: 20
 

docs/git-on-object-storage.md

@@ -51,13 +51,19 @@ to three stated axioms, each empirically gated per backend" does.
 
 ### v1 deployment architecture
 
-The implementation has *no per-repo persistent filesystem state*. Every request
+The implementation has *no authoritative per-repo filesystem state*. Every
-hydrates an ephemeral working tree from the published manifest, runs the
+request hydrates an ephemeral working tree from the published manifest, runs
-appropriate git subprocess against it, and drops the tree on scope exit:
+the appropriate git subprocess against it, and drops the tree on scope exit:
 read paths (`info/refs`, `upload-pack`) via `hydrate_for_read`, the write path
 (`receive-pack`) via `hydrate_for_write`, which also returns the `ParentState`
 the CAS at §Push step 7 predicates on. The relay is multi-instance-ready by
 construction: nothing on local disk needs to be coordinated between instances.
+Each process may retain a byte-bounded, process-lifetime cache of immutable
+pack/index pairs keyed by verified digest. Deployment mounts that cache on a
+per-pod ephemeral volume, so accounting and single-flight coordination stay
+local. Cache misses, restarts, and evictions only affect performance; object
+storage remains the source of truth, and per-request refs/HEAD are still
+materialized from the current manifest.
 
 The accepted v1 tradeoff: under concurrent same-repo pushes, every contender
 hydrates and runs receive-pack, and the CAS losers' subprocess work is
@@ -475,6 +481,7 @@ symbol search, not line counts.)
 | `GitStore::{put_pack, put_manifest, put_pointer}` (create-only + CAS) | `store.rs` |
 | `run_conformance_probe` (A1/A3 fail-closed startup gate) | `store.rs` + `main.rs` |
 | `hydrate_for_read` / `hydrate_for_write` | `hydrate.rs` |
+| Bounded digest-keyed pack/index cache and single-flight population | `pack_cache.rs` |
 | `ParentState { if_match, parent_digest, parent }` + `from_loaded`/`fresh` | `cas_publish.rs:154` |
 | `cas_publish(.., &parent_state) -> Result<CasSuccess, CasError>` | `cas_publish.rs:410` |
 | `CasError::Conflict { winner_manifest, winner_manifest_key }` (typed 412) | `cas_publish.rs:92` |