pageserver - OrderedBlockIter

* `wal_acceptor`: add HTTP handler, /metrics endpoint only, no authentication
* Two gauges are currently reported: `flush_lsn` and `commit_lsn`
* Add `DEFAULT_PG_LISTEN_PORT` and `DEFAULT_PG_LISTEN_PORT` consts for uniformity

.circleci/config.yml

@@ -24,6 +24,12 @@ jobs:
   # A job to build postgres
   build-postgres:
     executor: zenith-build-executor
+    parameters:
+      build_type:
+        type: enum
+        enum: ["debug", "release"]
+    environment:
+      BUILD_TYPE: << parameters.build_type >>
     steps:
         # Checkout the git repo (circleci doesn't have a flag to enable submodules here)
       - checkout
@@ -39,7 +45,7 @@ jobs:
           name: Restore postgres cache
           keys:
             # Restore ONLY if the rev key matches exactly
-            - v03-postgres-cache-{{ checksum "/tmp/cache-key-postgres" }}
+            - v04-postgres-cache-<< parameters.build_type >>-{{ checksum "/tmp/cache-key-postgres" }}
 
         # FIXME We could cache our own docker container, instead of installing packages every time.
       - run:
@@ -59,12 +65,12 @@ jobs:
             if [ ! -e tmp_install/bin/postgres ]; then
               # "depth 1" saves some time by not cloning the whole repo
               git submodule update --init --depth 1
-              make postgres
+              make postgres -j8
             fi
 
       - save_cache:
           name: Save postgres cache
-          key: v03-postgres-cache-{{ checksum "/tmp/cache-key-postgres" }}
+          key: v04-postgres-cache-<< parameters.build_type >>-{{ checksum "/tmp/cache-key-postgres" }}
           paths:
             - tmp_install
 
@@ -96,7 +102,7 @@ jobs:
           name: Restore postgres cache
           keys:
             # Restore ONLY if the rev key matches exactly
-            - v03-postgres-cache-{{ checksum "/tmp/cache-key-postgres" }}
+            - v04-postgres-cache-<< parameters.build_type >>-{{ checksum "/tmp/cache-key-postgres" }}
 
       - restore_cache:
           name: Restore rust cache
@@ -328,14 +334,18 @@ workflows:
   build_and_test:
     jobs:
       - check-codestyle
-      - build-postgres
+      - build-postgres:
+          name: build-postgres-<< matrix.build_type >>
+          matrix:
+            parameters:
+              build_type: ["debug", "release"]
       - build-zenith:
           name: build-zenith-<< matrix.build_type >>
           matrix:
             parameters:
               build_type: ["debug", "release"]
           requires:
-            - build-postgres
+            - build-postgres-<< matrix.build_type >>
       - run-pytest:
           name: pg_regress-tests-<< matrix.build_type >>
           matrix:

.dockerignore

@@ -11,4 +11,8 @@ test_output
 .vscode
 .zenith
 integration_tests/.zenith
-.mypy_cache
+.mypy_cache
+
+Dockerfile
+.dockerignore
+

Cargo.lock

@@ -2339,6 +2339,7 @@ dependencies = [
  "byteorder",
  "bytes",
  "clap",
+ "const_format",
  "crc32c",
  "daemonize",
  "fs2",
@@ -2358,6 +2359,7 @@ dependencies = [
  "tokio-stream",
  "walkdir",
  "workspace_hack",
+ "zenith_metrics",
  "zenith_utils",
 ]
 

Dockerfile

@@ -10,6 +10,7 @@ FROM zenithdb/build:buster AS pg-build
 WORKDIR /zenith
 COPY ./vendor/postgres vendor/postgres
 COPY ./Makefile Makefile
+ENV BUILD_TYPE release
 RUN make -j $(getconf _NPROCESSORS_ONLN) -s postgres
 RUN rm -rf postgres_install/build
 

Makefile

@@ -6,6 +6,18 @@ else
 	SECCOMP =
 endif
 
+#
+# We differentiate between release / debug build types using the BUILD_TYPE
+# environment variable.
+#
+ifeq ($(BUILD_TYPE),release)
+	PG_CONFIGURE_OPTS = --enable-debug
+	PG_CFLAGS = -O2 -g3 ${CFLAGS}
+else
+	PG_CONFIGURE_OPTS = --enable-debug --enable-cassert --enable-depend
+	PG_CFLAGS = -O0 -g3 ${CFLAGS}
+endif
+
 #
 # Top level Makefile to build Zenith and PostgreSQL
 #
@@ -30,10 +42,8 @@ tmp_install/build/config.status:
 	+@echo "Configuring postgres build"
 	mkdir -p tmp_install/build
 	(cd tmp_install/build && \
-	../../vendor/postgres/configure CFLAGS='-O0 -g3 $(CFLAGS)' \
-		--enable-cassert \
-		--enable-debug \
-		--enable-depend \
+	../../vendor/postgres/configure CFLAGS='$(PG_CFLAGS)' \
+		$(PG_CONFIGURE_OPTS) \
 		$(SECCOMP) \
 		--prefix=$(abspath tmp_install) > configure.log)
 

pageserver/src/branches.rs

@@ -419,7 +419,6 @@ fn create_timeline(
     let timelinedir = conf.timeline_path(&timelineid, tenantid);
 
     fs::create_dir(&timelinedir)?;
-    fs::create_dir(&timelinedir.join("wal"))?;
 
     if let Some(ancestor) = ancestor {
         let data = format!("{}@{}", ancestor.timelineid, ancestor.lsn);

pageserver/src/layered_repository.rs

@@ -23,6 +23,7 @@ use std::collections::hash_map::Entry;
 use std::collections::HashMap;
 use std::collections::{BTreeSet, HashSet};
 use std::convert::TryInto;
+use std::fs;
 use std::fs::{File, OpenOptions};
 use std::io::Write;
 use std::ops::Bound::Included;
@@ -30,7 +31,6 @@ use std::path::{Path, PathBuf};
 use std::sync::atomic::{AtomicUsize, Ordering};
 use std::sync::{Arc, Mutex, MutexGuard};
 use std::time::{Duration, Instant};
-use std::{fs, thread};
 
 use crate::layered_repository::inmemory_layer::FreezeLayers;
 use crate::relish::*;
@@ -57,6 +57,7 @@ mod image_layer;
 mod inmemory_layer;
 mod interval_tree;
 mod layer_map;
+mod page_versions;
 mod storage_layer;
 
 use delta_layer::DeltaLayer;
@@ -257,19 +258,6 @@ impl LayeredRepository {
                 timeline.init_current_logical_size()?;
 
                 let timeline = Arc::new(timeline);
-
-                // Load any new WAL after the last checkpoint into memory.
-                info!(
-                    "Loading WAL for timeline {} starting at {}",
-                    timelineid,
-                    timeline.get_last_record_lsn()
-                );
-
-                if cfg!(debug_assertions) {
-                    // check again after wal loading
-                    Self::assert_size_calculation_matches_offloaded(Arc::clone(&timeline));
-                }
-
                 timelines.insert(timelineid, timeline.clone());
                 Ok(timeline)
             }
@@ -535,24 +523,6 @@ impl LayeredRepository {
         totals.elapsed = now.elapsed();
         Ok(totals)
     }
-
-    fn assert_size_calculation_matches(incremental: usize, timeline: &LayeredTimeline) {
-        match timeline.get_current_logical_size_non_incremental(timeline.get_last_record_lsn()) {
-            Ok(non_incremental) => {
-                if incremental != non_incremental {
-                    error!("timeline size calculation diverged, incremental doesn't match non incremental. incremental={} non_incremental={}", incremental, non_incremental);
-                }
-            }
-            Err(e) => error!("failed to calculate non incremental timeline size: {:#}", e),
-        }
-    }
-
-    fn assert_size_calculation_matches_offloaded(timeline: Arc<LayeredTimeline>) {
-        let incremental = timeline.get_current_logical_size();
-        thread::spawn(move || {
-            Self::assert_size_calculation_matches(incremental, &timeline);
-        });
-    }
 }
 
 /// Metadata stored on disk for each timeline
@@ -1061,7 +1031,7 @@ impl LayeredTimeline {
             self.timelineid
         );
         let mut layers = self.layers.lock().unwrap();
-        let (imgfilenames, mut deltafilenames) =
+        let (imgfilenames, deltafilenames) =
             filename::list_files(self.conf, self.timelineid, self.tenantid)?;
 
         let timeline_path = self.conf.timeline_path(&self.timelineid, &self.tenantid);
@@ -1089,11 +1059,7 @@ impl LayeredTimeline {
             layers.insert_historic(Arc::new(layer));
         }
 
-        // Then for the Delta files. The delta files are created in order starting
-        // from the oldest file, because each DeltaLayer needs a reference to its
-        // predecessor.
-        deltafilenames.sort();
-
+        // Then for the Delta files.
         for filename in deltafilenames.iter() {
             ensure!(filename.start_lsn < filename.end_lsn);
             if filename.end_lsn > disk_consistent_lsn {
@@ -1106,27 +1072,12 @@ impl LayeredTimeline {
                 continue;
             }
 
-            let predecessor = layers.get(&filename.seg, filename.start_lsn);
-
-            let predecessor_str: String = if let Some(prec) = &predecessor {
-                prec.filename().display().to_string()
-            } else {
-                "none".to_string()
-            };
-
-            let layer = DeltaLayer::new(
-                self.conf,
-                self.timelineid,
-                self.tenantid,
-                filename,
-                predecessor,
-            );
+            let layer = DeltaLayer::new(self.conf, self.timelineid, self.tenantid, filename);
 
             info!(
-                "found layer {} on timeline {}, predecessor: {}",
+                "found layer {} on timeline {}",
                 layer.filename().display(),
                 self.timelineid,
-                predecessor_str,
             );
             layers.insert_historic(Arc::new(layer));
         }
@@ -1301,7 +1252,7 @@ impl LayeredTimeline {
             let start_lsn;
             if prev_layer.get_timeline_id() != self.timelineid {
                 // First modification on this timeline
-                start_lsn = self.ancestor_lsn;
+                start_lsn = self.ancestor_lsn + 1;
                 trace!(
                     "creating layer for write for {} at branch point {}/{}",
                     seg,
@@ -1453,19 +1404,6 @@ impl LayeredTimeline {
             // Finally, replace the frozen in-memory layer with the new on-disk layers
             layers.remove_historic(frozen.clone());
 
-            // If we created a successor InMemoryLayer, its predecessor is
-            // currently the frozen layer. We need to update the predecessor
-            // to be the latest on-disk layer.
-            if let Some(last_historic) = new_historics.last() {
-                if let Some(new_open) = &maybe_new_open {
-                    let maybe_old_predecessor =
-                        new_open.update_predecessor(Arc::clone(last_historic));
-                    let old_predecessor = maybe_old_predecessor
-                        .expect("new_open should always be a successor to frozen");
-                    assert!(layer_ptr_eq(frozen.as_ref(), old_predecessor.as_ref()));
-                }
-            }
-
             // Add the historics to the LayerMap
             for n in new_historics {
                 layers.insert_historic(n);
@@ -1772,12 +1710,20 @@ impl LayeredTimeline {
         loop {
             match layer_ref.get_page_reconstruct_data(blknum, curr_lsn, &mut data)? {
                 PageReconstructResult::Complete => break,
-                PageReconstructResult::Continue(cont_lsn, cont_layer) => {
+                PageReconstructResult::Continue(cont_lsn) => {
                     // Fetch base image / more WAL from the returned predecessor layer
-                    layer_arc = cont_layer;
-                    layer_ref = &*layer_arc;
-                    curr_lsn = cont_lsn;
-                    continue;
+                    if let Some((cont_layer, cont_lsn)) = self.get_layer_for_read(seg, cont_lsn)? {
+                        layer_arc = cont_layer;
+                        layer_ref = &*layer_arc;
+                        curr_lsn = cont_lsn;
+                        continue;
+                    } else {
+                        bail!(
+                            "could not find predecessor layer of segment {} at {}",
+                            seg.rel,
+                            cont_lsn
+                        );
+                    }
                 }
                 PageReconstructResult::Missing(lsn) => {
                     // Oops, we could not reconstruct the page.
@@ -1950,17 +1896,6 @@ pub fn dump_layerfile_from_path(path: &Path) -> Result<()> {
     Ok(())
 }
 
-/// Check for equality of Layer memory addresses
-fn layer_ptr_eq(l1: &dyn Layer, l2: &dyn Layer) -> bool {
-    let l1_ptr = l1 as *const dyn Layer;
-    let l2_ptr = l2 as *const dyn Layer;
-    // comparing *const dyn Layer will not only check for data address equality,
-    // but also for vtable address equality.
-    // to avoid this, we compare *const ().
-    // see here for more https://github.com/rust-lang/rust/issues/46139
-    std::ptr::eq(l1_ptr as *const (), l2_ptr as *const ())
-}
-
 /// Add a suffix to a layer file's name: .{num}.old
 /// Uses the first available num (starts at 0)
 fn rename_to_backup(path: PathBuf) -> anyhow::Result<()> {

pageserver/src/layered_repository/delta_layer.rs

@@ -45,10 +45,10 @@ use crate::layered_repository::storage_layer::{
 use crate::waldecoder;
 use crate::PageServerConf;
 use crate::{ZTenantId, ZTimelineId};
-use anyhow::{bail, Result};
+use anyhow::{bail, ensure, Result};
 use log::*;
 use serde::{Deserialize, Serialize};
-use std::collections::BTreeMap;
+use zenith_utils::vec_map::VecMap;
 // avoid binding to Write (conflicts with std::io::Write)
 // while being able to use std::fmt::Write's methods
 use std::fmt::Write as _;
@@ -57,7 +57,7 @@ use std::fs::File;
 use std::io::{BufWriter, Write};
 use std::ops::Bound::Included;
 use std::path::{Path, PathBuf};
-use std::sync::{Arc, Mutex, MutexGuard};
+use std::sync::{Mutex, MutexGuard};
 
 use bookfile::{Book, BookWriter};
 
@@ -65,6 +65,7 @@ use zenith_utils::bin_ser::BeSer;
 use zenith_utils::lsn::Lsn;
 
 use super::blob::{read_blob, BlobRange};
+use super::page_versions::OrderedBlockIter;
 
 // Magic constant to identify a Zenith delta file
 pub const DELTA_FILE_MAGIC: u32 = 0x5A616E01;
@@ -131,9 +132,6 @@ pub struct DeltaLayer {
 
     dropped: bool,
 
-    /// Predecessor layer
-    predecessor: Option<Arc<dyn Layer>>,
-
     inner: Mutex<DeltaLayerInner>,
 }
 
@@ -144,10 +142,10 @@ pub struct DeltaLayerInner {
 
     /// All versions of all pages in the file are are kept here.
     /// Indexed by block number and LSN.
-    page_version_metas: BTreeMap<(u32, Lsn), BlobRange>,
+    page_version_metas: VecMap<(u32, Lsn), BlobRange>,
 
     /// `relsizes` tracks the size of the relation at different points in time.
-    relsizes: BTreeMap<Lsn, u32>,
+    relsizes: VecMap<Lsn, u32>,
 }
 
 impl Layer for DeltaLayer {
@@ -218,10 +216,12 @@ impl Layer for DeltaLayer {
             // Scan the metadata BTreeMap backwards, starting from the given entry.
             let minkey = (blknum, Lsn(0));
             let maxkey = (blknum, lsn);
-            let mut iter = inner
+            let iter = inner
                 .page_version_metas
-                .range((Included(&minkey), Included(&maxkey)));
-            while let Some(((_blknum, _entry_lsn), blob_range)) = iter.next_back() {
+                .slice_range((Included(&minkey), Included(&maxkey)))
+                .iter()
+                .rev();
+            for ((_blknum, _lsn), blob_range) in iter {
                 let pv = PageVersion::des(&read_blob(&page_version_reader, blob_range)?)?;
 
                 if let Some(img) = pv.page_image {
@@ -247,16 +247,9 @@ impl Layer for DeltaLayer {
         }
 
         // If an older page image is needed to reconstruct the page, let the
-        // caller know about the predecessor layer.
+        // caller know.
         if need_image {
-            if let Some(cont_layer) = &self.predecessor {
-                Ok(PageReconstructResult::Continue(
-                    self.start_lsn,
-                    Arc::clone(cont_layer),
-                ))
-            } else {
-                Ok(PageReconstructResult::Missing(self.start_lsn))
-            }
+            Ok(PageReconstructResult::Continue(self.start_lsn))
         } else {
             Ok(PageReconstructResult::Complete)
         }
@@ -265,21 +258,22 @@ impl Layer for DeltaLayer {
     /// Get size of the relation at given LSN
     fn get_seg_size(&self, lsn: Lsn) -> Result<u32> {
         assert!(lsn >= self.start_lsn);
+        ensure!(
+            self.seg.rel.is_blocky(),
+            "get_seg_size() called on a non-blocky rel"
+        );
 
         // Scan the BTreeMap backwards, starting from the given entry.
         let inner = self.load()?;
-        let mut iter = inner.relsizes.range((Included(&Lsn(0)), Included(&lsn)));
+        let slice = inner
+            .relsizes
+            .slice_range((Included(&Lsn(0)), Included(&lsn)));
 
-        let result;
-        if let Some((_entry_lsn, entry)) = iter.next_back() {
-            result = *entry;
-        // Use the base image if needed
-        } else if let Some(predecessor) = &self.predecessor {
-            result = predecessor.get_seg_size(lsn)?;
+        if let Some((_entry_lsn, entry)) = slice.last() {
+            Ok(*entry)
         } else {
-            result = 0;
+            Err(anyhow::anyhow!("could not find seg size in delta layer"))
         }
-        Ok(result)
     }
 
     /// Does this segment exist at given LSN?
@@ -299,8 +293,8 @@ impl Layer for DeltaLayer {
     ///
     fn unload(&self) -> Result<()> {
         let mut inner = self.inner.lock().unwrap();
-        inner.page_version_metas = BTreeMap::new();
-        inner.relsizes = BTreeMap::new();
+        inner.page_version_metas = VecMap::default();
+        inner.relsizes = VecMap::default();
         inner.loaded = false;
         Ok(())
     }
@@ -326,13 +320,13 @@ impl Layer for DeltaLayer {
 
         println!("--- relsizes ---");
         let inner = self.load()?;
-        for (k, v) in inner.relsizes.iter() {
+        for (k, v) in inner.relsizes.as_slice() {
             println!("  {}: {}", k, v);
         }
         println!("--- page versions ---");
         let (_path, book) = self.open_book()?;
         let chapter = book.chapter_reader(PAGE_VERSIONS_CHAPTER)?;
-        for ((blk, lsn), blob_range) in inner.page_version_metas.iter() {
+        for ((blk, lsn), blob_range) in inner.page_version_metas.as_slice() {
             let mut desc = String::new();
 
             let buf = read_blob(&chapter, blob_range)?;
@@ -381,7 +375,7 @@ impl DeltaLayer {
     /// data structure with two btreemaps as we do, so passing the btreemaps is currently
     /// expedient.
     #[allow(clippy::too_many_arguments)]
-    pub fn create<'a>(
+    pub fn create(
         conf: &'static PageServerConf,
         timelineid: ZTimelineId,
         tenantid: ZTenantId,
@@ -389,10 +383,13 @@ impl DeltaLayer {
         start_lsn: Lsn,
         end_lsn: Lsn,
         dropped: bool,
-        predecessor: Option<Arc<dyn Layer>>,
-        page_versions: impl Iterator<Item = (&'a (u32, Lsn), &'a PageVersion)>,
-        relsizes: BTreeMap<Lsn, u32>,
+        page_versions: OrderedBlockIter,
+        relsizes: VecMap<Lsn, u32>,
     ) -> Result<DeltaLayer> {
+        if seg.rel.is_blocky() {
+            assert!(!relsizes.is_empty());
+        }
+
         let delta_layer = DeltaLayer {
             path_or_conf: PathOrConf::Conf(conf),
             timelineid,
@@ -403,10 +400,9 @@ impl DeltaLayer {
             dropped,
             inner: Mutex::new(DeltaLayerInner {
                 loaded: true,
-                page_version_metas: BTreeMap::new(),
+                page_version_metas: VecMap::default(),
                 relsizes,
             }),
-            predecessor,
         };
         let mut inner = delta_layer.inner.lock().unwrap();
 
@@ -423,26 +419,33 @@ impl DeltaLayer {
 
         let mut page_version_writer = BlobWriter::new(book, PAGE_VERSIONS_CHAPTER);
 
-        for (key, page_version) in page_versions {
-            let buf = PageVersion::ser(page_version)?;
-            let blob_range = page_version_writer.write_blob(&buf)?;
+        for (blknum, history) in page_versions {
+            for (lsn, page_version) in history.as_slice() {
+                if lsn >= &end_lsn {
+                    continue;
+                }
 
-            let old = inner.page_version_metas.insert(*key, blob_range);
+                let buf = PageVersion::ser(page_version)?;
+                let blob_range = page_version_writer.write_blob(&buf)?;
 
-            assert!(old.is_none());
+                inner
+                    .page_version_metas
+                    .append((blknum, *lsn), blob_range)
+                    .unwrap();
+            }
         }
 
         let book = page_version_writer.close()?;
 
         // Write out page versions
         let mut chapter = book.new_chapter(PAGE_VERSION_METAS_CHAPTER);
-        let buf = BTreeMap::ser(&inner.page_version_metas)?;
+        let buf = VecMap::ser(&inner.page_version_metas)?;
         chapter.write_all(&buf)?;
         let book = chapter.close()?;
 
         // and relsizes to separate chapter
         let mut chapter = book.new_chapter(REL_SIZES_CHAPTER);
-        let buf = BTreeMap::ser(&inner.relsizes)?;
+        let buf = VecMap::ser(&inner.relsizes)?;
         chapter.write_all(&buf)?;
         let book = chapter.close()?;
 
@@ -529,10 +532,10 @@ impl DeltaLayer {
         }
 
         let chapter = book.read_chapter(PAGE_VERSION_METAS_CHAPTER)?;
-        let page_version_metas = BTreeMap::des(&chapter)?;
+        let page_version_metas = VecMap::des(&chapter)?;
 
         let chapter = book.read_chapter(REL_SIZES_CHAPTER)?;
-        let relsizes = BTreeMap::des(&chapter)?;
+        let relsizes = VecMap::des(&chapter)?;
 
         debug!("loaded from {}", &path.display());
 
@@ -551,7 +554,6 @@ impl DeltaLayer {
         timelineid: ZTimelineId,
         tenantid: ZTenantId,
         filename: &DeltaFileName,
-        predecessor: Option<Arc<dyn Layer>>,
     ) -> DeltaLayer {
         DeltaLayer {
             path_or_conf: PathOrConf::Conf(conf),
@@ -563,10 +565,9 @@ impl DeltaLayer {
             dropped: filename.dropped,
             inner: Mutex::new(DeltaLayerInner {
                 loaded: false,
-                page_version_metas: BTreeMap::new(),
-                relsizes: BTreeMap::new(),
+                page_version_metas: VecMap::default(),
+                relsizes: VecMap::default(),
             }),
-            predecessor,
         }
     }
 
@@ -587,10 +588,9 @@ impl DeltaLayer {
             dropped: summary.dropped,
             inner: Mutex::new(DeltaLayerInner {
                 loaded: false,
-                page_version_metas: BTreeMap::new(),
-                relsizes: BTreeMap::new(),
+                page_version_metas: VecMap::default(),
+                relsizes: VecMap::default(),
             }),
-            predecessor: None,
         })
     }
 }

pageserver/src/layered_repository/filename.rs

@@ -290,11 +290,7 @@ pub fn list_files(
             deltafiles.push(deltafilename);
         } else if let Some(imgfilename) = ImageFileName::from_str(fname) {
             imgfiles.push(imgfilename);
-        } else if fname == "wal"
-            || fname == "metadata"
-            || fname == "ancestor"
-            || fname.ends_with(".old")
-        {
+        } else if fname == "metadata" || fname == "ancestor" || fname.ends_with(".old") {
             // ignore these
         } else {
             warn!("unrecognized filename in timeline dir: {}", fname);

pageserver/src/layered_repository/inmemory_layer.rs

@@ -12,18 +12,19 @@ use crate::layered_repository::{DeltaLayer, ImageLayer};
 use crate::repository::WALRecord;
 use crate::PageServerConf;
 use crate::{ZTenantId, ZTimelineId};
-use anyhow::{bail, Result};
+use anyhow::{bail, ensure, Result};
 use bytes::Bytes;
 use log::*;
 use std::cmp::min;
-use std::collections::BTreeMap;
-use std::ops::Bound::Included;
 use std::path::PathBuf;
 use std::sync::{Arc, RwLock};
+use zenith_utils::vec_map::VecMap;
 
 use zenith_utils::accum::Accum;
 use zenith_utils::lsn::Lsn;
 
+use super::page_versions::PageVersions;
+
 pub struct InMemoryLayer {
     conf: &'static PageServerConf,
     tenantid: ZTenantId,
@@ -45,6 +46,9 @@ pub struct InMemoryLayer {
     /// The above fields never change. The parts that do change are in 'inner',
     /// and protected by mutex.
     inner: RwLock<InMemoryLayerInner>,
+
+    /// Predecessor layer might be needed?
+    incremental: bool,
 }
 
 pub struct InMemoryLayerInner {
@@ -55,19 +59,20 @@ pub struct InMemoryLayerInner {
     /// All versions of all pages in the layer are are kept here.
     /// Indexed by block number and LSN.
     ///
-    page_versions: BTreeMap<(u32, Lsn), PageVersion>,
+    page_versions: PageVersions,
 
     ///
     /// `segsizes` tracks the size of the segment at different points in time.
     ///
-    segsizes: BTreeMap<Lsn, u32>,
+    /// For a blocky rel, there is always one entry, at the layer's start_lsn,
+    /// so that determining the size never depends on the predecessor layer. For
+    /// a non-blocky rel, 'segsizes' is not used and is always empty.
+    ///
+    segsizes: VecMap<Lsn, u32>,
 
     /// Writes are only allowed when true.
     /// Set to false when this layer is in the process of being replaced.
     writeable: bool,
-
-    /// Predecessor layer
-    predecessor: Option<Arc<dyn Layer>>,
 }
 
 impl InMemoryLayerInner {
@@ -81,12 +86,13 @@ impl InMemoryLayerInner {
 
     fn get_seg_size(&self, lsn: Lsn) -> u32 {
         // Scan the BTreeMap backwards, starting from the given entry.
-        let mut iter = self.segsizes.range((Included(&Lsn(0)), Included(&lsn)));
+        let slice = self.segsizes.slice_range(..=lsn);
 
-        if let Some((_entry_lsn, entry)) = iter.next_back() {
+        // We make sure there is always at least one entry
+        if let Some((_entry_lsn, entry)) = slice.last() {
             *entry
         } else {
-            0
+            panic!("could not find seg size in in-memory layer");
         }
     }
 }
@@ -164,18 +170,16 @@ impl Layer for InMemoryLayer {
 
         assert!(self.seg.blknum_in_seg(blknum));
 
-        let predecessor: Option<Arc<dyn Layer>>;
-
         {
             let inner = self.inner.read().unwrap();
 
-            // Scan the BTreeMap backwards, starting from reconstruct_data.lsn.
-            let minkey = (blknum, Lsn(0));
-            let maxkey = (blknum, lsn);
-            let mut iter = inner
+            // Scan the page versions backwards, starting from `lsn`.
+            let iter = inner
                 .page_versions
-                .range((Included(&minkey), Included(&maxkey)));
-            while let Some(((_blknum, _entry_lsn), entry)) = iter.next_back() {
+                .get_block_lsn_range(blknum, ..=lsn)
+                .iter()
+                .rev();
+            for (_entry_lsn, entry) in iter {
                 if let Some(img) = &entry.page_image {
                     reconstruct_data.page_img = Some(img.clone());
                     need_image = false;
@@ -192,16 +196,14 @@ impl Layer for InMemoryLayer {
                     bail!("no page image or WAL record for requested page");
                 }
             }
-
-            predecessor = inner.predecessor.clone();
             // release lock on 'inner'
         }
 
         // If an older page image is needed to reconstruct the page, let the
-        // caller know about the predecessor layer.
+        // caller know
         if need_image {
-            if let Some(cont_layer) = predecessor {
-                Ok(PageReconstructResult::Continue(self.start_lsn, cont_layer))
+            if self.incremental {
+                Ok(PageReconstructResult::Continue(Lsn(self.start_lsn.0 - 1)))
             } else {
                 Ok(PageReconstructResult::Missing(self.start_lsn))
             }
@@ -213,6 +215,10 @@ impl Layer for InMemoryLayer {
     /// Get size of the relation at given LSN
     fn get_seg_size(&self, lsn: Lsn) -> Result<u32> {
         assert!(lsn >= self.start_lsn);
+        ensure!(
+            self.seg.rel.is_blocky(),
+            "get_seg_size() called on a non-blocky rel"
+        );
 
         let inner = self.inner.read().unwrap();
         Ok(inner.get_seg_size(lsn))
@@ -252,8 +258,7 @@ impl Layer for InMemoryLayer {
     }
 
     fn is_incremental(&self) -> bool {
-        let inner = self.inner.read().unwrap();
-        inner.predecessor.is_some()
+        self.incremental
     }
 
     /// debugging function to print out the contents of the layer
@@ -271,18 +276,20 @@ impl Layer for InMemoryLayer {
             self.timelineid, self.seg, self.start_lsn, end_str
         );
 
-        for (k, v) in inner.segsizes.iter() {
+        for (k, v) in inner.segsizes.as_slice() {
             println!("segsizes {}: {}", k, v);
         }
 
-        for (k, v) in inner.page_versions.iter() {
-            println!(
-                "blk {} at {}: {}/{}\n",
-                k.0,
-                k.1,
-                v.page_image.is_some(),
-                v.record.is_some()
-            );
+        for (blknum, history) in inner.page_versions.ordered_block_iter() {
+            for (lsn, pv) in history.as_slice() {
+                println!(
+                    "blk {} at {}: {}/{}\n",
+                    blknum,
+                    lsn,
+                    pv.page_image.is_some(),
+                    pv.record.is_some()
+                );
+            }
         }
 
         Ok(())
@@ -332,6 +339,12 @@ impl InMemoryLayer {
             start_lsn
         );
 
+        // The segment is initially empty, so initialize 'segsizes' with 0.
+        let mut segsizes = VecMap::default();
+        if seg.rel.is_blocky() {
+            segsizes.append(start_lsn, 0).unwrap();
+        }
+
         Ok(InMemoryLayer {
             conf,
             timelineid,
@@ -340,12 +353,12 @@ impl InMemoryLayer {
             start_lsn,
             end_lsn: None,
             oldest_pending_lsn,
+            incremental: false,
             inner: RwLock::new(InMemoryLayerInner {
                 drop_lsn: None,
-                page_versions: BTreeMap::new(),
-                segsizes: BTreeMap::new(),
+                page_versions: PageVersions::default(),
+                segsizes,
                 writeable: true,
-                predecessor: None,
             }),
         })
     }
@@ -393,7 +406,7 @@ impl InMemoryLayer {
 
         inner.check_writeable()?;
 
-        let old = inner.page_versions.insert((blknum, lsn), pv);
+        let old = inner.page_versions.append_or_update_last(blknum, lsn, pv);
 
         if old.is_some() {
             // We already had an entry for this LSN. That's odd..
@@ -438,7 +451,9 @@ impl InMemoryLayer {
                         gapblknum,
                         blknum
                     );
-                    let old = inner.page_versions.insert((gapblknum, lsn), zeropv);
+                    let old = inner
+                        .page_versions
+                        .append_or_update_last(gapblknum, lsn, zeropv);
                     // We already had an entry for this LSN. That's odd..
 
                     if old.is_some() {
@@ -449,7 +464,7 @@ impl InMemoryLayer {
                     }
                 }
 
-                inner.segsizes.insert(lsn, newsize);
+                inner.segsizes.append_or_update_last(lsn, newsize).unwrap();
                 return Ok(newsize - oldsize);
             }
         }
@@ -458,6 +473,10 @@ impl InMemoryLayer {
 
     /// Remember that the relation was truncated at given LSN
     pub fn put_truncation(&self, lsn: Lsn, segsize: u32) -> WriteResult<()> {
+        assert!(
+            self.seg.rel.is_blocky(),
+            "put_truncation() called on a non-blocky rel"
+        );
         self.assert_not_frozen();
 
         let mut inner = self.inner.write().unwrap();
@@ -467,7 +486,7 @@ impl InMemoryLayer {
         let oldsize = inner.get_seg_size(lsn);
         assert!(segsize < oldsize);
 
-        let old = inner.segsizes.insert(lsn, segsize);
+        let old = inner.segsizes.append_or_update_last(lsn, segsize).unwrap();
 
         if old.is_some() {
             // We already had an entry for this LSN. That's odd..
@@ -518,11 +537,11 @@ impl InMemoryLayer {
             start_lsn,
         );
 
-        // For convenience, copy the segment size from the predecessor layer
-        let mut segsizes = BTreeMap::new();
+        // Copy the segment size at the start LSN from the predecessor layer.
+        let mut segsizes = VecMap::default();
         if seg.rel.is_blocky() {
             let size = src.get_seg_size(start_lsn)?;
-            segsizes.insert(start_lsn, size);
+            segsizes.append(start_lsn, size).unwrap();
         }
 
         Ok(InMemoryLayer {
@@ -533,12 +552,12 @@ impl InMemoryLayer {
             start_lsn,
             end_lsn: None,
             oldest_pending_lsn,
+            incremental: true,
             inner: RwLock::new(InMemoryLayerInner {
                 drop_lsn: None,
-                page_versions: BTreeMap::new(),
+                page_versions: PageVersions::default(),
                 segsizes,
                 writeable: true,
-                predecessor: Some(src),
             }),
         })
     }
@@ -589,28 +608,18 @@ impl InMemoryLayer {
 
         // Divide all the page versions into old and new
         // at the 'cutoff_lsn' point.
-        let mut before_segsizes = BTreeMap::new();
-        let mut after_segsizes = BTreeMap::new();
         let mut after_oldest_lsn: Accum<Lsn> = Accum(None);
-        for (lsn, size) in inner.segsizes.iter() {
-            if *lsn > cutoff_lsn {
-                after_segsizes.insert(*lsn, *size);
-                after_oldest_lsn.accum(min, *lsn);
-            } else {
-                before_segsizes.insert(*lsn, *size);
-            }
+
+        let cutoff_lsn_exclusive = Lsn(cutoff_lsn.0 + 1);
+
+        let (before_segsizes, mut after_segsizes) = inner.segsizes.split_at(&cutoff_lsn_exclusive);
+        if let Some((lsn, _size)) = after_segsizes.as_slice().first() {
+            after_oldest_lsn.accum(min, *lsn);
         }
 
-        let mut before_page_versions = BTreeMap::new();
-        let mut after_page_versions = BTreeMap::new();
-        for ((blknum, lsn), pv) in inner.page_versions.iter() {
-            if *lsn > cutoff_lsn {
-                after_page_versions.insert((*blknum, *lsn), pv.clone());
-                after_oldest_lsn.accum(min, *lsn);
-            } else {
-                before_page_versions.insert((*blknum, *lsn), pv.clone());
-            }
-        }
+        let (before_page_versions, after_page_versions) = inner
+            .page_versions
+            .split_at(cutoff_lsn_exclusive, &mut after_oldest_lsn);
 
         let frozen = Arc::new(InMemoryLayer {
             conf: self.conf,
@@ -620,12 +629,12 @@ impl InMemoryLayer {
             start_lsn: self.start_lsn,
             end_lsn: Some(cutoff_lsn),
             oldest_pending_lsn: self.start_lsn,
+            incremental: self.incremental,
             inner: RwLock::new(InMemoryLayerInner {
                 drop_lsn: inner.drop_lsn,
                 page_versions: before_page_versions,
                 segsizes: before_segsizes,
                 writeable: false,
-                predecessor: inner.predecessor.clone(),
             }),
         });
 
@@ -640,8 +649,11 @@ impl InMemoryLayer {
             )?;
 
             let new_inner = new_open.inner.get_mut().unwrap();
-            new_inner.page_versions.append(&mut after_page_versions);
-            new_inner.segsizes.append(&mut after_segsizes);
+            // Ensure page_versions doesn't contain anything
+            // so we can just replace it
+            assert!(new_inner.page_versions.is_empty());
+            new_inner.page_versions = after_page_versions;
+            new_inner.segsizes.extend(&mut after_segsizes).unwrap();
 
             Some(Arc::new(new_open))
         } else {
@@ -679,8 +691,6 @@ impl InMemoryLayer {
         let inner = self.inner.read().unwrap();
         assert!(!inner.writeable);
 
-        let predecessor = inner.predecessor.clone();
-
         if let Some(drop_lsn) = inner.drop_lsn {
             let delta_layer = DeltaLayer::create(
                 self.conf,
@@ -690,8 +700,7 @@ impl InMemoryLayer {
                 self.start_lsn,
                 drop_lsn,
                 true,
-                predecessor,
-                inner.page_versions.iter(),
+                inner.page_versions.ordered_block_iter(),
                 inner.segsizes.clone(),
             )?;
             trace!(
@@ -705,21 +714,11 @@ impl InMemoryLayer {
 
         let end_lsn = self.end_lsn.unwrap();
 
-        let mut before_segsizes = BTreeMap::new();
-        for (lsn, size) in inner.segsizes.iter() {
-            if *lsn <= end_lsn {
-                before_segsizes.insert(*lsn, *size);
-            }
-        }
-        let mut before_page_versions = inner.page_versions.iter().filter(|tup| {
-            let ((_blknum, lsn), _pv) = tup;
-
-            *lsn < end_lsn
-        });
-
         let mut frozen_layers: Vec<Arc<dyn Layer>> = Vec::new();
 
         if self.start_lsn != end_lsn {
+            let (before_segsizes, _after_segsizes) = inner.segsizes.split_at(&Lsn(end_lsn.0 + 1));
+
             // Write the page versions before the cutoff to disk.
             let delta_layer = DeltaLayer::create(
                 self.conf,
@@ -729,8 +728,7 @@ impl InMemoryLayer {
                 self.start_lsn,
                 end_lsn,
                 false,
-                predecessor,
-                before_page_versions,
+                inner.page_versions.ordered_block_iter(),
                 before_segsizes,
             )?;
             frozen_layers.push(Arc::new(delta_layer));
@@ -741,7 +739,10 @@ impl InMemoryLayer {
                 end_lsn
             );
         } else {
-            assert!(before_page_versions.next().is_none());
+            for (_blknum, history) in inner.page_versions.ordered_block_iter() {
+                let (lsn, _pv) = history.as_slice().first().unwrap();
+                assert!(lsn >= &end_lsn);
+            }
         }
 
         drop(inner);
@@ -753,9 +754,4 @@ impl InMemoryLayer {
 
         Ok(frozen_layers)
     }
-
-    pub fn update_predecessor(&self, predecessor: Arc<dyn Layer>) -> Option<Arc<dyn Layer>> {
-        let mut inner = self.inner.write().unwrap();
-        inner.predecessor.replace(predecessor)
-    }
 }

pageserver/src/layered_repository/page_versions.rs

@@ -0,0 +1,132 @@
+use std::{collections::HashMap, ops::RangeBounds};
+
+use zenith_utils::{accum::Accum, lsn::Lsn, vec_map::VecMap};
+
+use super::storage_layer::PageVersion;
+
+const EMPTY_SLICE: &[(Lsn, PageVersion)] = &[];
+
+#[derive(Debug, Default)]
+pub struct PageVersions(HashMap<u32, VecMap<Lsn, PageVersion>>);
+
+impl PageVersions {
+    pub fn is_empty(&self) -> bool {
+        self.0.is_empty()
+    }
+
+    pub fn append_or_update_last(
+        &mut self,
+        blknum: u32,
+        lsn: Lsn,
+        page_version: PageVersion,
+    ) -> Option<PageVersion> {
+        let map = self.0.entry(blknum).or_insert_with(VecMap::default);
+        map.append_or_update_last(lsn, page_version).unwrap()
+    }
+
+    /// Get a range of [`PageVersions`] in a block
+    pub fn get_block_lsn_range<R: RangeBounds<Lsn>>(
+        &self,
+        blknum: u32,
+        range: R,
+    ) -> &[(Lsn, PageVersion)] {
+        self.0
+            .get(&blknum)
+            .map(|vec_map| vec_map.slice_range(range))
+            .unwrap_or(EMPTY_SLICE)
+    }
+
+    /// Split the page version map into two.
+    ///
+    /// Left contains everything up to and not including [`cutoff_lsn`].
+    /// Right contains [`cutoff_lsn`] and everything after.
+    pub fn split_at(&self, cutoff_lsn: Lsn, after_oldest_lsn: &mut Accum<Lsn>) -> (Self, Self) {
+        let mut before_blocks = HashMap::new();
+        let mut after_blocks = HashMap::new();
+
+        for (blknum, vec_map) in self.0.iter() {
+            let (before_versions, after_versions) = vec_map.split_at(&cutoff_lsn);
+
+            if !before_versions.is_empty() {
+                let old = before_blocks.insert(*blknum, before_versions);
+                assert!(old.is_none());
+            }
+
+            if !after_versions.is_empty() {
+                let (first_lsn, _first_pv) = &after_versions.as_slice()[0];
+                after_oldest_lsn.accum(std::cmp::min, *first_lsn);
+
+                let old = after_blocks.insert(*blknum, after_versions);
+                assert!(old.is_none());
+            }
+        }
+
+        (Self(before_blocks), Self(after_blocks))
+    }
+
+    /// Iterate through block-history pairs in block order.
+    pub fn ordered_block_iter(&self) -> OrderedBlockIter<'_> {
+        let mut ordered_blocks: Vec<u32> = self.0.keys().cloned().collect();
+        ordered_blocks.sort_unstable();
+
+        OrderedBlockIter {
+            page_versions: self,
+            ordered_blocks,
+            cur_block_idx: 0,
+        }
+    }
+}
+
+pub struct OrderedBlockIter<'a> {
+    page_versions: &'a PageVersions,
+
+    ordered_blocks: Vec<u32>,
+    cur_block_idx: usize,
+}
+
+impl<'a> Iterator for OrderedBlockIter<'a> {
+    type Item = (u32, &'a VecMap<Lsn, PageVersion>);
+
+    fn next(&mut self) -> Option<Self::Item> {
+        let blknum: u32 = *self.ordered_blocks.get(self.cur_block_idx)?;
+        self.cur_block_idx += 1;
+        Some((blknum, self.page_versions.0.get(&blknum).unwrap()))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    const EMPTY_PAGE_VERSION: PageVersion = PageVersion {
+        page_image: None,
+        record: None,
+    };
+
+    #[test]
+    fn test_ordered_iter() {
+        let mut page_versions = PageVersions::default();
+        const BLOCKS: u32 = 1000;
+        const LSNS: u64 = 50;
+
+        for blknum in 0..BLOCKS {
+            for lsn in 0..LSNS {
+                let old = page_versions.append_or_update_last(blknum, Lsn(lsn), EMPTY_PAGE_VERSION);
+                assert!(old.is_none());
+            }
+        }
+
+        let mut iter = page_versions.ordered_block_iter();
+        for blknum in 0..BLOCKS {
+            let (actual_blknum, vec_map) = iter.next().unwrap();
+            let slice = vec_map.as_slice();
+            assert_eq!(actual_blknum, blknum);
+            assert_eq!(slice.len(), LSNS as usize);
+            for lsn in 0..LSNS {
+                assert_eq!(Lsn(lsn), slice[lsn as usize].0);
+            }
+        }
+        assert!(iter.next().is_none());
+        assert!(iter.next().is_none()); // should be robust against excessive next() calls
+    }
+}

pageserver/src/layered_repository/storage_layer.rs

@@ -10,7 +10,6 @@ use bytes::Bytes;
 use serde::{Deserialize, Serialize};
 use std::fmt;
 use std::path::PathBuf;
-use std::sync::Arc;
 
 use zenith_utils::lsn::Lsn;
 
@@ -87,9 +86,9 @@ pub struct PageReconstructData {
 pub enum PageReconstructResult {
     /// Got all the data needed to reconstruct the requested page
     Complete,
-    /// This layer didn't contain all the required data, the caller should collect
-    /// more data from the returned predecessor layer at the returned LSN.
-    Continue(Lsn, Arc<dyn Layer>),
+    /// This layer didn't contain all the required data, the caller should look up
+    /// the predecessor layer at the returned LSN and collect more data from there.
+    Continue(Lsn),
     /// This layer didn't contain data needed to reconstruct the page version at
     /// the returned LSN. This is usually considered an error, but might be OK
     /// in some circumstances.
@@ -145,8 +144,8 @@ pub trait Layer: Send + Sync {
     ///
     /// See PageReconstructResult for possible return values. The collected data
     /// is appended to reconstruct_data; the caller should pass an empty struct
-    /// on first call. If this returns PageReconstructResult::Continue, call
-    /// again on the returned predecessor layer with the same 'reconstruct_data'
+    /// on first call. If this returns PageReconstructResult::Continue, look up
+    /// the predecessor layer and call again with the same 'reconstruct_data'
     /// to collect more data.
     fn get_page_reconstruct_data(
         &self,

pageserver/src/lib.rs

@@ -126,10 +126,6 @@ impl PageServerConf {
         self.timeline_path(timelineid, tenantid).join("ancestor")
     }
 
-    fn wal_dir_path(&self, timelineid: &ZTimelineId, tenantid: &ZTenantId) -> PathBuf {
-        self.timeline_path(timelineid, tenantid).join("wal")
-    }
-
     //
     // Postgres distribution paths
     //

pageserver/src/tenant_mgr.rs

@@ -100,10 +100,6 @@ pub fn create_repository_for_tenant(
     Ok(())
 }
 
-pub fn insert_repository_for_tenant(tenantid: ZTenantId, repo: Arc<dyn Repository>) {
-    access_repository().insert(tenantid, repo);
-}
-
 pub fn get_repository_for_tenant(tenantid: ZTenantId) -> Result<Arc<dyn Repository>> {
     access_repository()
         .get(&tenantid)

pageserver/src/walreceiver.rs

@@ -16,14 +16,11 @@ use log::*;
 use postgres::fallible_iterator::FallibleIterator;
 use postgres::replication::ReplicationIter;
 use postgres::{Client, NoTls, SimpleQueryMessage, SimpleQueryRow};
-use postgres_ffi::xlog_utils::*;
 use postgres_ffi::*;
 use postgres_protocol::message::backend::ReplicationMessage;
 use postgres_types::PgLsn;
 use std::cell::Cell;
-use std::cmp::{max, min};
 use std::collections::HashMap;
-use std::fs;
 use std::str::FromStr;
 use std::sync::Mutex;
 use std::thread;
@@ -125,7 +122,7 @@ fn thread_main(conf: &'static PageServerConf, timelineid: ZTimelineId, tenantid:
 }
 
 fn walreceiver_main(
-    conf: &PageServerConf,
+    _conf: &PageServerConf,
     timelineid: ZTimelineId,
     wal_producer_connstr: &str,
     tenantid: ZTenantId,
@@ -195,7 +192,6 @@ fn walreceiver_main(
                 let data = xlog_data.data();
                 let startlsn = Lsn::from(xlog_data.wal_start());
                 let endlsn = startlsn + data.len() as u64;
-                let prev_last_rec_lsn = last_rec_lsn;
 
                 trace!("received XLogData between {} and {}", startlsn, endlsn);
 
@@ -236,34 +232,6 @@ fn walreceiver_main(
                     last_rec_lsn = lsn;
                 }
 
-                // Somewhat arbitrarily, if we have at least 10 complete wal segments (16 MB each),
-                // "checkpoint" the repository to flush all the changes from WAL we've processed
-                // so far to disk. After this, we don't need the original WAL anymore, and it
-                // can be removed. This is probably too aggressive for production, but it's useful
-                // to expose bugs now.
-                //
-                // TODO: We don't actually dare to remove the WAL. It's useful for debugging,
-                // and we might it for logical decoding other things in the future. Although
-                // we should also be able to fetch it back from the WAL safekeepers or S3 if
-                // needed.
-                if prev_last_rec_lsn.segment_number(pg_constants::WAL_SEGMENT_SIZE)
-                    != last_rec_lsn.segment_number(pg_constants::WAL_SEGMENT_SIZE)
-                {
-                    info!("switched segment {} to {}", prev_last_rec_lsn, last_rec_lsn);
-                    let (oldest_segno, newest_segno) = find_wal_file_range(
-                        conf,
-                        &timelineid,
-                        pg_constants::WAL_SEGMENT_SIZE,
-                        last_rec_lsn,
-                        &tenantid,
-                    )?;
-
-                    if newest_segno - oldest_segno >= 10 {
-                        // TODO: This is where we could remove WAL older than last_rec_lsn.
-                        //remove_wal_files(timelineid, pg_constants::WAL_SEGMENT_SIZE, last_rec_lsn)?;
-                    }
-                }
-
                 if !caught_up && endlsn >= end_of_wal {
                     info!("caught up at LSN {}", endlsn);
                     caught_up = true;
@@ -285,7 +253,7 @@ fn walreceiver_main(
                 );
 
                 if reply_requested {
-                    Some(timeline.get_last_record_lsn())
+                    Some(last_rec_lsn)
                 } else {
                     None
                 }
@@ -309,47 +277,6 @@ fn walreceiver_main(
     Ok(())
 }
 
-fn find_wal_file_range(
-    conf: &PageServerConf,
-    timeline: &ZTimelineId,
-    wal_seg_size: usize,
-    written_upto: Lsn,
-    tenant: &ZTenantId,
-) -> Result<(u64, u64)> {
-    let written_upto_segno = written_upto.segment_number(wal_seg_size);
-
-    let mut oldest_segno = written_upto_segno;
-    let mut newest_segno = written_upto_segno;
-    // Scan the wal directory, and count how many WAL filed we could remove
-    let wal_dir = conf.wal_dir_path(timeline, tenant);
-    for entry in fs::read_dir(wal_dir)? {
-        let entry = entry?;
-        let path = entry.path();
-
-        if path.is_dir() {
-            continue;
-        }
-
-        let filename = path.file_name().unwrap().to_str().unwrap();
-
-        if IsXLogFileName(filename) {
-            let (segno, _tli) = XLogFromFileName(filename, wal_seg_size);
-
-            if segno > written_upto_segno {
-                // that's strange.
-                warn!("there is a WAL file from future at {}", path.display());
-                continue;
-            }
-
-            oldest_segno = min(oldest_segno, segno);
-            newest_segno = max(newest_segno, segno);
-        }
-    }
-    // FIXME: would be good to assert that there are no gaps in the WAL files
-
-    Ok((oldest_segno, newest_segno))
-}
-
 /// Data returned from the postgres `IDENTIFY_SYSTEM` command
 ///
 /// See the [postgres docs] for more details.

pageserver/src/walredo.rs

@@ -17,11 +17,9 @@
 //! records. It achieves it by dropping privileges before replaying
 //! any WAL records, so that even if an attacker hijacks the Postgres
 //! process, he cannot escape out of it.
-
-mod nonrel;
-mod request;
-
-use bytes::Bytes;
+//!
+use byteorder::{ByteOrder, LittleEndian};
+use bytes::{Buf, BufMut, Bytes, BytesMut};
 use lazy_static::lazy_static;
 use log::*;
 use serde::Serialize;
@@ -39,12 +37,21 @@ use tokio::io::{AsyncReadExt, AsyncWriteExt};
 use tokio::process::{ChildStdin, ChildStdout, Command};
 use tokio::time::timeout;
 use zenith_metrics::{register_histogram, register_int_counter, Histogram, IntCounter};
+use zenith_utils::bin_ser::BeSer;
 use zenith_utils::lsn::Lsn;
 use zenith_utils::zid::ZTenantId;
 
 use crate::relish::*;
 use crate::repository::WALRecord;
+use crate::waldecoder::XlMultiXactCreate;
+use crate::waldecoder::XlXactParsedRecord;
 use crate::PageServerConf;
+use postgres_ffi::nonrelfile_utils::mx_offset_to_flags_bitshift;
+use postgres_ffi::nonrelfile_utils::mx_offset_to_flags_offset;
+use postgres_ffi::nonrelfile_utils::mx_offset_to_member_offset;
+use postgres_ffi::nonrelfile_utils::transaction_id_set_status;
+use postgres_ffi::pg_constants;
+use postgres_ffi::XLogRecord;
 
 ///
 /// `RelTag` + block number (`blknum`) gives us a unique id of the page in the cluster.
@@ -85,7 +92,7 @@ pub trait WalRedoManager: Send + Sync {
 /// a Repository object without launching the real WAL redo process.
 ///
 pub struct DummyRedoManager {}
-impl WalRedoManager for DummyRedoManager {
+impl crate::walredo::WalRedoManager for DummyRedoManager {
     fn request_redo(
         &self,
         _rel: RelishTag,
@@ -98,7 +105,7 @@ impl WalRedoManager for DummyRedoManager {
     }
 }
 
-const TIMEOUT: Duration = Duration::from_secs(20);
+static TIMEOUT: Duration = Duration::from_secs(20);
 
 // Metrics collected on WAL redo operations
 //
@@ -242,20 +249,166 @@ impl PostgresRedoManager {
         process: &mut PostgresRedoProcess,
         request: &WalRedoRequest,
     ) -> Result<Bytes, WalRedoError> {
+        let rel = request.rel;
+        let blknum = request.blknum;
+        let lsn = request.lsn;
+        let base_img = request.base_img.clone();
+        let records = &request.records;
+
+        let nrecords = records.len();
+
         let start = Instant::now();
 
-        let apply_result = if let RelishTag::Relation(rel) = request.rel {
+        let apply_result: Result<Bytes, Error>;
+        if let RelishTag::Relation(rel) = rel {
             // Relational WAL records are applied using wal-redo-postgres
-            let buf_tag = BufferTag {
-                rel,
-                blknum: request.blknum,
-            };
-            process
-                .apply_wal_records(buf_tag, &request.base_img, &request.records)
-                .await
+            let buf_tag = BufferTag { rel, blknum };
+            apply_result = process.apply_wal_records(buf_tag, base_img, records).await;
         } else {
-            Ok(nonrel::apply_nonrel(request))
-        };
+            // Non-relational WAL records are handled here, with custom code that has the
+            // same effects as the corresponding Postgres WAL redo function.
+            const ZERO_PAGE: [u8; 8192] = [0u8; 8192];
+            let mut page = BytesMut::new();
+            if let Some(fpi) = base_img {
+                // If full-page image is provided, then use it...
+                page.extend_from_slice(&fpi[..]);
+            } else {
+                // otherwise initialize page with zeros
+                page.extend_from_slice(&ZERO_PAGE);
+            }
+            // Apply all collected WAL records
+            for record in records {
+                let mut buf = record.rec.clone();
+
+                WAL_REDO_RECORD_COUNTER.inc();
+
+                // 1. Parse XLogRecord struct
+                // FIXME: refactor to avoid code duplication.
+                let xlogrec = XLogRecord::from_bytes(&mut buf);
+
+                //move to main data
+                // TODO probably, we should store some records in our special format
+                // to avoid this weird parsing on replay
+                let skip = (record.main_data_offset - pg_constants::SIZEOF_XLOGRECORD) as usize;
+                if buf.remaining() > skip {
+                    buf.advance(skip);
+                }
+
+                if xlogrec.xl_rmid == pg_constants::RM_XACT_ID {
+                    // Transaction manager stuff
+                    let rec_segno = match rel {
+                        RelishTag::Slru { slru, segno } => {
+                            assert!(
+                                slru == SlruKind::Clog,
+                                "Not valid XACT relish tag {:?}",
+                                rel
+                            );
+                            segno
+                        }
+                        _ => panic!("Not valid XACT relish tag {:?}", rel),
+                    };
+                    let parsed_xact =
+                        XlXactParsedRecord::decode(&mut buf, xlogrec.xl_xid, xlogrec.xl_info);
+                    if parsed_xact.info == pg_constants::XLOG_XACT_COMMIT
+                        || parsed_xact.info == pg_constants::XLOG_XACT_COMMIT_PREPARED
+                    {
+                        transaction_id_set_status(
+                            parsed_xact.xid,
+                            pg_constants::TRANSACTION_STATUS_COMMITTED,
+                            &mut page,
+                        );
+                        for subxact in &parsed_xact.subxacts {
+                            let pageno = *subxact as u32 / pg_constants::CLOG_XACTS_PER_PAGE;
+                            let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
+                            let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
+                            // only update xids on the requested page
+                            if rec_segno == segno && blknum == rpageno {
+                                transaction_id_set_status(
+                                    *subxact,
+                                    pg_constants::TRANSACTION_STATUS_COMMITTED,
+                                    &mut page,
+                                );
+                            }
+                        }
+                    } else if parsed_xact.info == pg_constants::XLOG_XACT_ABORT
+                        || parsed_xact.info == pg_constants::XLOG_XACT_ABORT_PREPARED
+                    {
+                        transaction_id_set_status(
+                            parsed_xact.xid,
+                            pg_constants::TRANSACTION_STATUS_ABORTED,
+                            &mut page,
+                        );
+                        for subxact in &parsed_xact.subxacts {
+                            let pageno = *subxact as u32 / pg_constants::CLOG_XACTS_PER_PAGE;
+                            let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
+                            let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
+                            // only update xids on the requested page
+                            if rec_segno == segno && blknum == rpageno {
+                                transaction_id_set_status(
+                                    *subxact,
+                                    pg_constants::TRANSACTION_STATUS_ABORTED,
+                                    &mut page,
+                                );
+                            }
+                        }
+                    }
+                } else if xlogrec.xl_rmid == pg_constants::RM_MULTIXACT_ID {
+                    // Multixact operations
+                    let info = xlogrec.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
+                    if info == pg_constants::XLOG_MULTIXACT_CREATE_ID {
+                        let xlrec = XlMultiXactCreate::decode(&mut buf);
+                        if let RelishTag::Slru {
+                            slru,
+                            segno: rec_segno,
+                        } = rel
+                        {
+                            if slru == SlruKind::MultiXactMembers {
+                                for i in 0..xlrec.nmembers {
+                                    let pageno =
+                                        i / pg_constants::MULTIXACT_MEMBERS_PER_PAGE as u32;
+                                    let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
+                                    let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
+                                    if segno == rec_segno && rpageno == blknum {
+                                        // update only target block
+                                        let offset = xlrec.moff + i;
+                                        let memberoff = mx_offset_to_member_offset(offset);
+                                        let flagsoff = mx_offset_to_flags_offset(offset);
+                                        let bshift = mx_offset_to_flags_bitshift(offset);
+                                        let mut flagsval =
+                                            LittleEndian::read_u32(&page[flagsoff..flagsoff + 4]);
+                                        flagsval &= !(((1
+                                            << pg_constants::MXACT_MEMBER_BITS_PER_XACT)
+                                            - 1)
+                                            << bshift);
+                                        flagsval |= xlrec.members[i as usize].status << bshift;
+                                        LittleEndian::write_u32(
+                                            &mut page[flagsoff..flagsoff + 4],
+                                            flagsval,
+                                        );
+                                        LittleEndian::write_u32(
+                                            &mut page[memberoff..memberoff + 4],
+                                            xlrec.members[i as usize].xid,
+                                        );
+                                    }
+                                }
+                            } else {
+                                // Multixact offsets SLRU
+                                let offs = (xlrec.mid
+                                    % pg_constants::MULTIXACT_OFFSETS_PER_PAGE as u32
+                                    * 4) as usize;
+                                LittleEndian::write_u32(&mut page[offs..offs + 4], xlrec.moff);
+                            }
+                        } else {
+                            panic!();
+                        }
+                    } else {
+                        panic!();
+                    }
+                }
+            }
+
+            apply_result = Ok::<Bytes, Error>(page.freeze());
+        }
 
         let duration = start.elapsed();
 
@@ -263,9 +416,9 @@ impl PostgresRedoManager {
 
         debug!(
             "applied {} WAL records in {} ms to reconstruct page image at LSN {}",
-            request.records.len(),
+            nrecords,
             duration.as_millis(),
-            request.lsn
+            lsn
         );
 
         if let Err(e) = apply_result {
@@ -390,11 +543,12 @@ impl PostgresRedoProcess {
     async fn apply_wal_records(
         &mut self,
         tag: BufferTag,
-        base_img: &Option<Bytes>,
+        base_img: Option<Bytes>,
         records: &[WALRecord],
     ) -> Result<Bytes, std::io::Error> {
-        let stdin = &mut self.stdin;
         let stdout = &mut self.stdout;
+        // Buffer the writes to avoid a lot of small syscalls.
+        let mut stdin = tokio::io::BufWriter::new(&mut self.stdin);
 
         // We do three things simultaneously: send the old base image and WAL records to
         // the child process's stdin, read the result from child's stdout, and forward any logging
@@ -404,24 +558,140 @@ impl PostgresRedoProcess {
         // 'f_stdout' below reads the result back. And 'f_stderr', which was spawned into the
         // tokio runtime in the 'launch' function already, forwards the logging.
         let f_stdin = async {
-            let buf = request::serialize_request(tag, base_img, records);
+            // Send base image, if any. (If the record initializes the page, previous page
+            // version is not needed.)
+            timeout(
+                TIMEOUT,
+                stdin.write_all(&build_begin_redo_for_block_msg(tag)),
+            )
+            .await??;
+            if base_img.is_some() {
+                timeout(
+                    TIMEOUT,
+                    stdin.write_all(&build_push_page_msg(tag, base_img.unwrap())),
+                )
+                .await??;
+            }
 
-            timeout(TIMEOUT, stdin.write_all(&buf)).await??;
+            // Send WAL records.
+            for rec in records.iter() {
+                let r = rec.clone();
+
+                WAL_REDO_RECORD_COUNTER.inc();
+
+                stdin
+                    .write_all(&build_apply_record_msg(r.lsn, r.rec))
+                    .await?;
+
+                //debug!("sent WAL record to wal redo postgres process ({:X}/{:X}",
+                //       r.lsn >> 32, r.lsn & 0xffff_ffff);
+            }
+            //debug!("sent {} WAL records to wal redo postgres process ({:X}/{:X}",
+            //       records.len(), lsn >> 32, lsn & 0xffff_ffff);
+
+            // Send GetPage command to get the result back
+            timeout(TIMEOUT, stdin.write_all(&build_get_page_msg(tag))).await??;
             timeout(TIMEOUT, stdin.flush()).await??;
-
+            //debug!("sent GetPage for {}", tag.blknum);
             Ok::<(), Error>(())
         };
 
         // Read back new page image
         let f_stdout = async {
-            let mut buf = vec![0u8; 8192];
+            let mut buf = [0u8; 8192];
 
             timeout(TIMEOUT, stdout.read_exact(&mut buf)).await??;
             //debug!("got response for {}", tag.blknum);
-            Ok::<Vec<u8>, Error>(buf)
+            Ok::<[u8; 8192], Error>(buf)
         };
 
-        let (buf, _) = tokio::try_join!(f_stdout, f_stdin)?;
-        Ok::<Bytes, Error>(Bytes::from(buf))
+        let res = tokio::try_join!(f_stdout, f_stdin)?;
+
+        let buf = res.0;
+
+        Ok::<Bytes, Error>(Bytes::from(std::vec::Vec::from(buf)))
     }
 }
+
+// Functions for constructing messages to send to the postgres WAL redo
+// process. See vendor/postgres/src/backend/tcop/zenith_wal_redo.c for
+// explanation of the protocol.
+
+fn build_begin_redo_for_block_msg(tag: BufferTag) -> Bytes {
+    let len = 4 + 1 + 4 * 4;
+    let mut buf = BytesMut::with_capacity(1 + len);
+
+    buf.put_u8(b'B');
+    buf.put_u32(len as u32);
+
+    // FIXME: this is a temporary hack that should go away when we refactor
+    // the postgres protocol serialization + handlers.
+    //
+    // BytesMut is a dynamic growable buffer, used a lot in tokio code but
+    // not in the std library. To write to a BytesMut from a serde serializer,
+    // we need to either:
+    // - pre-allocate the required buffer space. This is annoying because we
+    //   shouldn't care what the exact serialized size is-- that's the
+    //   serializer's job.
+    // - Or, we need to create a temporary "writer" (which implements the
+    //   `Write` trait). It's a bit awkward, because the writer consumes the
+    //   underlying BytesMut, and we need to extract it later with
+    //   `into_inner`.
+    let mut writer = buf.writer();
+    tag.ser_into(&mut writer)
+        .expect("serialize BufferTag should always succeed");
+    let buf = writer.into_inner();
+
+    debug_assert!(buf.len() == 1 + len);
+
+    buf.freeze()
+}
+
+fn build_push_page_msg(tag: BufferTag, base_img: Bytes) -> Bytes {
+    assert!(base_img.len() == 8192);
+
+    let len = 4 + 1 + 4 * 4 + base_img.len();
+    let mut buf = BytesMut::with_capacity(1 + len);
+
+    buf.put_u8(b'P');
+    buf.put_u32(len as u32);
+    let mut writer = buf.writer();
+    tag.ser_into(&mut writer)
+        .expect("serialize BufferTag should always succeed");
+    let mut buf = writer.into_inner();
+    buf.put(base_img);
+
+    debug_assert!(buf.len() == 1 + len);
+
+    buf.freeze()
+}
+
+fn build_apply_record_msg(endlsn: Lsn, rec: Bytes) -> Bytes {
+    let len = 4 + 8 + rec.len();
+    let mut buf = BytesMut::with_capacity(1 + len);
+
+    buf.put_u8(b'A');
+    buf.put_u32(len as u32);
+    buf.put_u64(endlsn.0);
+    buf.put(rec);
+
+    debug_assert!(buf.len() == 1 + len);
+
+    buf.freeze()
+}
+
+fn build_get_page_msg(tag: BufferTag) -> Bytes {
+    let len = 4 + 1 + 4 * 4;
+    let mut buf = BytesMut::with_capacity(1 + len);
+
+    buf.put_u8(b'G');
+    buf.put_u32(len as u32);
+    let mut writer = buf.writer();
+    tag.ser_into(&mut writer)
+        .expect("serialize BufferTag should always succeed");
+    let buf = writer.into_inner();
+
+    debug_assert!(buf.len() == 1 + len);
+
+    buf.freeze()
+}

pageserver/src/walredo/nonrel.rs

@@ -1,161 +0,0 @@
-use byteorder::{ByteOrder, LittleEndian};
-use bytes::{Buf, Bytes, BytesMut};
-use postgres_ffi::{
-    nonrelfile_utils::{
-        mx_offset_to_flags_bitshift, mx_offset_to_flags_offset, mx_offset_to_member_offset,
-        transaction_id_set_status,
-    },
-    pg_constants, XLogRecord,
-};
-
-use crate::{
-    relish::{RelishTag, SlruKind},
-    waldecoder::{XlMultiXactCreate, XlXactParsedRecord},
-};
-
-use super::WalRedoRequest;
-
-pub(super) fn apply_nonrel(request: &WalRedoRequest) -> Bytes {
-    let rel = request.rel;
-    let blknum = request.blknum;
-
-    // Non-relational WAL records are handled here, with custom code that has the
-    // same effects as the corresponding Postgres WAL redo function.
-    const ZERO_PAGE: [u8; 8192] = [0u8; 8192];
-    let mut page = BytesMut::new();
-    if let Some(fpi) = &request.base_img {
-        // If full-page image is provided, then use it...
-        page.extend_from_slice(&fpi[..]);
-    } else {
-        // otherwise initialize page with zeros
-        page.extend_from_slice(&ZERO_PAGE);
-    }
-    // Apply all collected WAL records
-    for record in &request.records {
-        let mut buf = record.rec.clone();
-
-        super::WAL_REDO_RECORD_COUNTER.inc();
-
-        // 1. Parse XLogRecord struct
-        // FIXME: refactor to avoid code duplication.
-        let xlogrec = XLogRecord::from_bytes(&mut buf);
-
-        //move to main data
-        // TODO probably, we should store some records in our special format
-        // to avoid this weird parsing on replay
-        let skip = (record.main_data_offset - pg_constants::SIZEOF_XLOGRECORD) as usize;
-        if buf.remaining() > skip {
-            buf.advance(skip);
-        }
-
-        if xlogrec.xl_rmid == pg_constants::RM_XACT_ID {
-            // Transaction manager stuff
-            let rec_segno = match rel {
-                RelishTag::Slru { slru, segno } => {
-                    assert!(
-                        slru == SlruKind::Clog,
-                        "Not valid XACT relish tag {:?}",
-                        rel
-                    );
-                    segno
-                }
-                _ => panic!("Not valid XACT relish tag {:?}", rel),
-            };
-            let parsed_xact = XlXactParsedRecord::decode(&mut buf, xlogrec.xl_xid, xlogrec.xl_info);
-            if parsed_xact.info == pg_constants::XLOG_XACT_COMMIT
-                || parsed_xact.info == pg_constants::XLOG_XACT_COMMIT_PREPARED
-            {
-                transaction_id_set_status(
-                    parsed_xact.xid,
-                    pg_constants::TRANSACTION_STATUS_COMMITTED,
-                    &mut page,
-                );
-                for subxact in &parsed_xact.subxacts {
-                    let pageno = *subxact as u32 / pg_constants::CLOG_XACTS_PER_PAGE;
-                    let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
-                    let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
-                    // only update xids on the requested page
-                    if rec_segno == segno && blknum == rpageno {
-                        transaction_id_set_status(
-                            *subxact,
-                            pg_constants::TRANSACTION_STATUS_COMMITTED,
-                            &mut page,
-                        );
-                    }
-                }
-            } else if parsed_xact.info == pg_constants::XLOG_XACT_ABORT
-                || parsed_xact.info == pg_constants::XLOG_XACT_ABORT_PREPARED
-            {
-                transaction_id_set_status(
-                    parsed_xact.xid,
-                    pg_constants::TRANSACTION_STATUS_ABORTED,
-                    &mut page,
-                );
-                for subxact in &parsed_xact.subxacts {
-                    let pageno = *subxact as u32 / pg_constants::CLOG_XACTS_PER_PAGE;
-                    let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
-                    let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
-                    // only update xids on the requested page
-                    if rec_segno == segno && blknum == rpageno {
-                        transaction_id_set_status(
-                            *subxact,
-                            pg_constants::TRANSACTION_STATUS_ABORTED,
-                            &mut page,
-                        );
-                    }
-                }
-            }
-        } else if xlogrec.xl_rmid == pg_constants::RM_MULTIXACT_ID {
-            // Multixact operations
-            let info = xlogrec.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
-            if info == pg_constants::XLOG_MULTIXACT_CREATE_ID {
-                let xlrec = XlMultiXactCreate::decode(&mut buf);
-                if let RelishTag::Slru {
-                    slru,
-                    segno: rec_segno,
-                } = rel
-                {
-                    if slru == SlruKind::MultiXactMembers {
-                        for i in 0..xlrec.nmembers {
-                            let pageno = i / pg_constants::MULTIXACT_MEMBERS_PER_PAGE as u32;
-                            let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
-                            let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
-                            if segno == rec_segno && rpageno == blknum {
-                                // update only target block
-                                let offset = xlrec.moff + i;
-                                let memberoff = mx_offset_to_member_offset(offset);
-                                let flagsoff = mx_offset_to_flags_offset(offset);
-                                let bshift = mx_offset_to_flags_bitshift(offset);
-                                let mut flagsval =
-                                    LittleEndian::read_u32(&page[flagsoff..flagsoff + 4]);
-                                flagsval &= !(((1 << pg_constants::MXACT_MEMBER_BITS_PER_XACT)
-                                    - 1)
-                                    << bshift);
-                                flagsval |= xlrec.members[i as usize].status << bshift;
-                                LittleEndian::write_u32(
-                                    &mut page[flagsoff..flagsoff + 4],
-                                    flagsval,
-                                );
-                                LittleEndian::write_u32(
-                                    &mut page[memberoff..memberoff + 4],
-                                    xlrec.members[i as usize].xid,
-                                );
-                            }
-                        }
-                    } else {
-                        // Multixact offsets SLRU
-                        let offs = (xlrec.mid % pg_constants::MULTIXACT_OFFSETS_PER_PAGE as u32 * 4)
-                            as usize;
-                        LittleEndian::write_u32(&mut page[offs..offs + 4], xlrec.moff);
-                    }
-                } else {
-                    panic!();
-                }
-            } else {
-                panic!();
-            }
-        }
-    }
-
-    page.freeze()
-}

pageserver/src/walredo/request.rs

@@ -1,86 +0,0 @@
-///! Functions for constructing messages to send to the postgres WAL redo
-///! process. See vendor/postgres/src/backend/tcop/zenith_wal_redo.c for
-///! explanation of the protocol.
-use bytes::{BufMut, Bytes, BytesMut};
-use zenith_utils::{bin_ser::BeSer, lsn::Lsn};
-
-use crate::repository::WALRecord;
-
-use super::BufferTag;
-
-pub fn serialize_request(
-    tag: BufferTag,
-    base_img: &Option<Bytes>,
-    records: &[WALRecord],
-) -> BytesMut {
-    let mut capacity = 1 + BEGIN_REDO_MSG_LEN;
-    if base_img.is_some() {
-        capacity += 1 + PUSH_PAGE_MSG_LEN;
-    }
-    capacity += (1 + APPLY_MSG_HEADER_LEN) * records.len();
-    capacity += records.iter().map(|rec| rec.rec.len()).sum::<usize>();
-    capacity += 1 + GET_PAGE_MSG_LEN;
-
-    let mut buf = BytesMut::with_capacity(capacity);
-
-    build_begin_redo_for_block_msg(&mut buf, tag);
-
-    if let Some(base_img) = base_img.as_ref() {
-        build_push_page_msg(&mut buf, tag, base_img);
-    }
-
-    for record in records {
-        build_apply_record_msg(&mut buf, record.lsn, &record.rec);
-    }
-
-    build_get_page_msg(&mut buf, tag);
-
-    debug_assert_eq!(capacity, buf.len());
-
-    buf
-}
-
-const TAG_LEN: usize = 4 * 4;
-const PAGE_SIZE: usize = 8192;
-const BEGIN_REDO_MSG_LEN: usize = 4 + 1 + TAG_LEN;
-const PUSH_PAGE_MSG_LEN: usize = 4 + 1 + TAG_LEN + PAGE_SIZE;
-const APPLY_MSG_HEADER_LEN: usize = 4 + 8;
-const GET_PAGE_MSG_LEN: usize = 4 + 1 + TAG_LEN;
-
-fn build_begin_redo_for_block_msg(buf: &mut BytesMut, tag: BufferTag) {
-    buf.put_u8(b'B');
-    buf.put_u32(BEGIN_REDO_MSG_LEN as u32);
-
-    // TODO tag is serialized multiple times
-    // let's try to serialize it just once
-    // or make the protocol less repetitive
-    tag.ser_into(&mut buf.writer())
-        .expect("serialize BufferTag should always succeed");
-}
-
-fn build_push_page_msg(buf: &mut BytesMut, tag: BufferTag, base_img: &Bytes) {
-    debug_assert_eq!(base_img.len(), PAGE_SIZE);
-
-    buf.put_u8(b'P');
-    buf.put_u32(PUSH_PAGE_MSG_LEN as u32);
-    tag.ser_into(&mut buf.writer())
-        .expect("serialize BufferTag should always succeed");
-    buf.extend(base_img);
-}
-
-fn build_apply_record_msg(buf: &mut BytesMut, endlsn: Lsn, rec: &Bytes) {
-    buf.put_u8(b'A');
-
-    let len = APPLY_MSG_HEADER_LEN + rec.len();
-    buf.put_u32(len as u32);
-
-    buf.put_u64(endlsn.0);
-    buf.extend(rec);
-}
-
-fn build_get_page_msg(buf: &mut BytesMut, tag: BufferTag) {
-    buf.put_u8(b'G');
-    buf.put_u32(GET_PAGE_MSG_LEN as u32);
-    tag.ser_into(&mut buf.writer())
-        .expect("serialize BufferTag should always succeed");
-}

test_runner/batch_others/test_branch_behind.py

@@ -33,11 +33,11 @@ def test_branch_behind(zenith_cli, pageserver: ZenithPageserver, postgres: Postg
     main_cur.execute('''
         INSERT INTO foo
             SELECT 'long string to consume some space' || g
-            FROM generate_series(1, 100000) g
+            FROM generate_series(1, 200000) g
     ''')
     main_cur.execute('SELECT pg_current_wal_insert_lsn()')
     lsn_b = main_cur.fetchone()[0]
-    print('LSN after 100100 rows: ' + lsn_b)
+    print('LSN after 200100 rows: ' + lsn_b)
 
     # Branch at the point where only 100 rows were inserted
     zenith_cli.run(["branch", "test_branch_behind_hundred", "test_branch_behind@" + lsn_a])
@@ -46,15 +46,15 @@ def test_branch_behind(zenith_cli, pageserver: ZenithPageserver, postgres: Postg
     main_cur.execute('''
         INSERT INTO foo
             SELECT 'long string to consume some space' || g
-            FROM generate_series(1, 100000) g
+            FROM generate_series(1, 200000) g
     ''')
     main_cur.execute('SELECT pg_current_wal_insert_lsn()')
 
     main_cur.execute('SELECT pg_current_wal_insert_lsn()')
     lsn_c = main_cur.fetchone()[0]
-    print('LSN after 200100 rows: ' + lsn_c)
+    print('LSN after 400100 rows: ' + lsn_c)
 
-    # Branch at the point where only 200 rows were inserted
+    # Branch at the point where only 200100 rows were inserted
     zenith_cli.run(["branch", "test_branch_behind_more", "test_branch_behind@" + lsn_b])
 
     pg_hundred = postgres.create_start("test_branch_behind_hundred")
@@ -70,11 +70,11 @@ def test_branch_behind(zenith_cli, pageserver: ZenithPageserver, postgres: Postg
     more_pg_conn = pg_more.connect()
     more_cur = more_pg_conn.cursor()
     more_cur.execute('SELECT count(*) FROM foo')
-    assert more_cur.fetchone() == (100100, )
+    assert more_cur.fetchone() == (200100, )
 
     # All the rows are visible on the main branch
     main_cur.execute('SELECT count(*) FROM foo')
-    assert main_cur.fetchone() == (200100, )
+    assert main_cur.fetchone() == (400100, )
 
     # Check bad lsn's for branching
 

test_runner/fixtures/benchmark_fixture.py

@@ -140,6 +140,30 @@ class ZenithBenchmarker:
                             re.MULTILINE)
         return int(round(float(matches.group(1))))
 
+    def get_peak_mem(self, pageserver) -> int:
+        """
+        Fetch the "maxrss" metric from the pageserver
+        """
+        # Fetch all the exposed prometheus metrics from page server
+        all_metrics = pageserver.http_client().get_metrics()
+        # See comment in get_io_writes()
+        matches = re.search(r'^pageserver_maxrss_kb (\S+)$', all_metrics,
+                            re.MULTILINE)
+        return int(round(float(matches.group(1))))
+
+    def get_timeline_size(self, repo_dir: str, tenantid: str, timelineid: str):
+        """
+        Calculate the on-disk size of a timeline
+        """
+        path = "{}/tenants/{}/timelines/{}".format(repo_dir, tenantid, timelineid)
+
+        totalbytes = 0
+        for root, dirs, files in os.walk(path):
+            for name in files:
+                totalbytes += os.path.getsize(os.path.join(root, name))
+
+        return totalbytes
+
     @contextmanager
     def record_pageserver_writes(self, pageserver, metric_name):
         """

test_runner/fixtures/zenith_fixtures.py

@@ -794,12 +794,17 @@ def read_pid(path: Path):
     return int(path.read_text())
 
 
+@dataclass
+class WalAcceptorPort:
+    pg: int
+    http: int
+
 @dataclass
 class WalAcceptor:
     """ An object representing a running wal acceptor daemon. """
     wa_bin_path: Path
     data_dir: Path
-    port: int
+    port: WalAcceptorPort
     num: int # identifier for logging
     pageserver_port: int
     auth_token: Optional[str] = None
@@ -811,7 +816,8 @@ class WalAcceptor:
 
         cmd = [str(self.wa_bin_path)]
         cmd.extend(["-D", str(self.data_dir)])
-        cmd.extend(["-l", f"localhost:{self.port}"])
+        cmd.extend(["--listen-pg", f"localhost:{self.port.pg}"])
+        cmd.extend(["--listen-http", f"localhost:{self.port.http}"])
         cmd.append("--daemonize")
         cmd.append("--no-sync")
         # Tell page server it can receive WAL from this WAL safekeeper
@@ -868,14 +874,14 @@ class WalAcceptor:
 
         # "replication=0" hacks psycopg not to send additional queries
         # on startup, see https://github.com/psycopg/psycopg2/pull/482
-        connstr = f"host=localhost port={self.port} replication=0 options='-c ztimelineid={timeline_id} ztenantid={tenant_id}'"
+        connstr = f"host=localhost port={self.port.pg} replication=0 options='-c ztimelineid={timeline_id} ztenantid={tenant_id}'"
 
         with closing(psycopg2.connect(connstr)) as conn:
             # server doesn't support transactions
             conn.autocommit = True
             with conn.cursor() as cur:
                 request_json = json.dumps(request)
-                print(f"JSON_CTRL request on port {self.port}: {request_json}")
+                print(f"JSON_CTRL request on port {self.port.pg}: {request_json}")
                 cur.execute("JSON_CTRL " + request_json)
                 all = cur.fetchall()
                 print(f"JSON_CTRL response: {all[0][0]}")
@@ -898,7 +904,10 @@ class WalAcceptorFactory:
         wa = WalAcceptor(
             wa_bin_path=self.wa_bin_path,
             data_dir=self.data_dir / "wal_acceptor_{}".format(wa_num),
-            port=self.port_distributor.get_port(),
+            port=WalAcceptorPort(
+                pg=self.port_distributor.get_port(),
+                http=self.port_distributor.get_port(),
+            ),
             num=wa_num,
             pageserver_port=self.pageserver_port,
             auth_token=auth_token,
@@ -922,7 +931,7 @@ class WalAcceptorFactory:
 
     def get_connstrs(self) -> str:
         """ Get list of wal acceptor endpoints suitable for wal_acceptors GUC  """
-        return ','.join(["localhost:{}".format(wa.port) for wa in self.instances])
+        return ','.join(["localhost:{}".format(wa.port.pg) for wa in self.instances])
 
 
 @zenfixture

test_runner/performance/test_bulk_insert.py

@@ -4,19 +4,6 @@ from fixtures.zenith_fixtures import PostgresFactory, ZenithPageserver
 
 pytest_plugins = ("fixtures.zenith_fixtures", "fixtures.benchmark_fixture")
 
-def get_timeline_size(repo_dir: str, tenantid: str, timelineid: str):
-    path = "{}/tenants/{}/timelines/{}".format(repo_dir, tenantid, timelineid)
-
-    totalbytes = 0
-    for root, dirs, files in os.walk(path):
-        for name in files:
-            totalbytes += os.path.getsize(os.path.join(root, name))
-
-        if 'wal' in dirs:
-            dirs.remove('wal')  # don't visit 'wal' subdirectory
-
-    return totalbytes
-
 #
 # Run bulk INSERT test.
 #
@@ -25,6 +12,7 @@ def get_timeline_size(repo_dir: str, tenantid: str, timelineid: str):
 # 1. Time to INSERT 5 million rows
 # 2. Disk writes
 # 3. Disk space used
+# 4. Peak memory usage
 #
 def test_bulk_insert(postgres: PostgresFactory, pageserver: ZenithPageserver, pg_bin, zenith_cli, zenbenchmark, repo_dir: str):
     # Create a branch for us
@@ -55,6 +43,9 @@ def test_bulk_insert(postgres: PostgresFactory, pageserver: ZenithPageserver, pg
                     # time and I/O
                     pscur.execute(f"do_gc {pageserver.initial_tenant} {timeline} 0")
 
+            # Record peak memory usage
+            zenbenchmark.record("peak_mem", zenbenchmark.get_peak_mem(pageserver) / 1024, 'MB')
+
             # Report disk space used by the repository
-            timeline_size = get_timeline_size(repo_dir, pageserver.initial_tenant, timeline)
+            timeline_size = zenbenchmark.get_timeline_size(repo_dir, pageserver.initial_tenant, timeline)
             zenbenchmark.record('size', timeline_size / (1024*1024), 'MB')

test_runner/performance/test_perf_pgbench.py

@@ -4,19 +4,6 @@ from fixtures.zenith_fixtures import PostgresFactory, ZenithPageserver
 
 pytest_plugins = ("fixtures.zenith_fixtures", "fixtures.benchmark_fixture")
 
-def get_timeline_size(repo_dir: str, tenantid: str, timelineid: str):
-    path = "{}/tenants/{}/timelines/{}".format(repo_dir, tenantid, timelineid)
-
-    totalbytes = 0
-    for root, dirs, files in os.walk(path):
-        for name in files:
-            totalbytes += os.path.getsize(os.path.join(root, name))
-
-        if 'wal' in dirs:
-            dirs.remove('wal')  # don't visit 'wal' subdirectory
-
-    return totalbytes
-
 #
 # Run a very short pgbench test.
 #
@@ -64,5 +51,5 @@ def test_pgbench(postgres: PostgresFactory, pageserver: ZenithPageserver, pg_bin
     pscur.execute(f"do_gc {pageserver.initial_tenant} {timeline} 0")
 
     # Report disk space used by the repository
-    timeline_size = get_timeline_size(repo_dir, pageserver.initial_tenant, timeline)
+    timeline_size = zenbenchmark.get_timeline_size(repo_dir, pageserver.initial_tenant, timeline)
     zenbenchmark.record('size', timeline_size / (1024*1024), 'MB')

test_runner/performance/test_write_amplification.py

@@ -0,0 +1,74 @@
+# Demonstrate Write Amplification with naive oldest-first layer checkpointing
+# algorithm.
+#
+# In each iteration of the test, we create a new table that's slightly under 10
+# MB in size (10 MB is the current "segment size" used by the page server). Then
+# we make a tiny update to all the tables already created. This creates a WAL
+# pattern where you have a lot of updates on one segment (the newly created
+# one), alternating with a small updates on all relations. This is the worst
+# case scenario for the naive checkpointing policy where we write out the layers
+# in LSN order, writing the oldest layer first. That creates a new 10 MB image
+# layer to be created for each of those small updates.  This is the Write
+# Amplification problem at its finest.
+import os
+from contextlib import closing
+from fixtures.zenith_fixtures import PostgresFactory, ZenithPageserver
+
+pytest_plugins = ("fixtures.zenith_fixtures", "fixtures.benchmark_fixture")
+
+def test_write_amplification(postgres: PostgresFactory, pageserver: ZenithPageserver, pg_bin, zenith_cli, zenbenchmark, repo_dir: str):
+    # Create a branch for us
+    zenith_cli.run(["branch", "test_write_amplification", "empty"])
+
+    pg = postgres.create_start('test_write_amplification')
+    print("postgres is running on 'test_write_amplification' branch")
+
+    # Open a connection directly to the page server that we'll use to force
+    # flushing the layers to disk
+    psconn = pageserver.connect();
+    pscur = psconn.cursor()
+
+    with closing(pg.connect()) as conn:
+        with conn.cursor() as cur:
+            # Get the timeline ID of our branch. We need it for the 'do_gc' command
+            cur.execute("SHOW zenith.zenith_timeline")
+            timeline = cur.fetchone()[0]
+
+            with zenbenchmark.record_pageserver_writes(pageserver, 'pageserver_writes'):
+                with zenbenchmark.record_duration('run'):
+
+                    # NOTE: Because each iteration updates every table already created,
+                    # the runtime and write amplification is O(n^2), where n is the
+                    # number of iterations.
+                    for i in range(25):
+                        cur.execute(f'''
+                        CREATE TABLE tbl{i} AS
+                            SELECT g as i, 'long string to consume some space' || g as t
+                            FROM generate_series(1, 100000) g
+                        ''')
+                        cur.execute(f"create index on tbl{i} (i);")
+                        for j in range(1, i):
+                            cur.execute(f"delete from tbl{j} where i = {i}")
+
+                        # Force checkpointing. As of this writing, we don't have
+                        # a back-pressure mechanism, and the page server cannot
+                        # keep up digesting and checkpointing the WAL at the
+                        # rate that it is generated. If we don't force a
+                        # checkpoint, the WAL will just accumulate in memory
+                        # until you hit OOM error. So in effect, we use much
+                        # more memory to hold the incoming WAL, and write them
+                        # out in larger batches than we'd really want. Using
+                        # more memory hides the write amplification problem this
+                        # test tries to demonstrate.
+                        #
+                        # The write amplification problem is real, and using
+                        # more memory isn't the right solution. We could
+                        # demonstrate the effect also by generating the WAL
+                        # slower, adding some delays in this loop.  But forcing
+                        # the the checkpointing and GC makes the test go faster,
+                        # with the same total I/O effect.
+                        pscur.execute(f"do_gc {pageserver.initial_tenant} {timeline} 0")
+
+            # Report disk space used by the repository
+            timeline_size = zenbenchmark.get_timeline_size(repo_dir, pageserver.initial_tenant, timeline)
+            zenbenchmark.record('size', timeline_size / (1024*1024), 'MB')

walkeeper/Cargo.toml

@@ -28,9 +28,11 @@ humantime = "2.1.0"
 walkdir = "2"
 serde = { version = "1.0", features = ["derive"] }
 hex = "0.4.3"
+const_format = "0.2.21"
 
 # FIXME: 'pageserver' is needed for ZTimelineId. Refactor
 pageserver = { path = "../pageserver" }
 postgres_ffi = { path = "../postgres_ffi" }
 workspace_hack = { path = "../workspace_hack" }
+zenith_metrics = { path = "../zenith_metrics" }
 zenith_utils = { path = "../zenith_utils" }

walkeeper/src/bin/wal_acceptor.rs

@@ -3,18 +3,23 @@
 //
 use anyhow::Result;
 use clap::{App, Arg};
+use const_format::formatcp;
 use daemonize::Daemonize;
 use log::*;
 use std::env;
+use std::net::TcpListener;
 use std::path::{Path, PathBuf};
 use std::thread;
+use zenith_utils::http::endpoint;
 use zenith_utils::logging;
 
+use walkeeper::defaults::{DEFAULT_HTTP_LISTEN_ADDR, DEFAULT_PG_LISTEN_ADDR};
 use walkeeper::s3_offload;
 use walkeeper::wal_service;
 use walkeeper::WalAcceptorConf;
 
 fn main() -> Result<()> {
+    zenith_metrics::set_common_metrics_prefix("safekeeper");
     let arg_matches = App::new("Zenith wal_acceptor")
         .about("Store WAL stream to local file system and push it to WAL receivers")
         .arg(
@@ -25,11 +30,18 @@ fn main() -> Result<()> {
                 .help("Path to the WAL acceptor data directory"),
         )
         .arg(
-            Arg::with_name("listen")
+            Arg::with_name("listen-pg")
                 .short("l")
-                .long("listen")
+                .long("listen-pg")
+                .alias("listen") // for compatibility
                 .takes_value(true)
-                .help("listen for incoming connections on ip:port (default: 127.0.0.1:5454)"),
+                .help(formatcp!("listen for incoming WAL data connections on ip:port (default: {DEFAULT_PG_LISTEN_ADDR})")),
+        )
+        .arg(
+            Arg::with_name("listen-http")
+                .long("listen-http")
+                .takes_value(true)
+                .help(formatcp!("http endpoint address for metrics on ip:port (default: {DEFAULT_HTTP_LISTEN_ADDR})")),
         )
         .arg(
             Arg::with_name("pageserver")
@@ -70,7 +82,8 @@ fn main() -> Result<()> {
         daemonize: false,
         no_sync: false,
         pageserver_addr: None,
-        listen_addr: "localhost:5454".to_string(),
+        listen_pg_addr: DEFAULT_PG_LISTEN_ADDR.to_string(),
+        listen_http_addr: DEFAULT_HTTP_LISTEN_ADDR.to_string(),
         ttl: None,
         recall_period: None,
         pageserver_auth_token: env::var("PAGESERVER_AUTH_TOKEN").ok(),
@@ -91,8 +104,12 @@ fn main() -> Result<()> {
         conf.daemonize = true;
     }
 
-    if let Some(addr) = arg_matches.value_of("listen") {
-        conf.listen_addr = addr.to_owned();
+    if let Some(addr) = arg_matches.value_of("listen-pg") {
+        conf.listen_pg_addr = addr.to_owned();
+    }
+
+    if let Some(addr) = arg_matches.value_of("listen-http") {
+        conf.listen_http_addr = addr.to_owned();
     }
 
     if let Some(addr) = arg_matches.value_of("pageserver") {
@@ -114,6 +131,11 @@ fn start_wal_acceptor(conf: WalAcceptorConf) -> Result<()> {
     let log_filename = conf.data_dir.join("wal_acceptor.log");
     let (_scope_guard, log_file) = logging::init(log_filename, conf.daemonize)?;
 
+    let http_listener = TcpListener::bind(conf.listen_http_addr.clone()).map_err(|e| {
+        error!("failed to bind to address {}: {}", conf.listen_http_addr, e);
+        e
+    })?;
+
     if conf.daemonize {
         info!("daemonizing...");
 
@@ -136,6 +158,16 @@ fn start_wal_acceptor(conf: WalAcceptorConf) -> Result<()> {
 
     let mut threads = Vec::new();
 
+    let http_endpoint_thread = thread::Builder::new()
+        .name("http_endpoint_thread".into())
+        .spawn(move || {
+            // No authentication at all: read-only metrics only, early stage.
+            let router = endpoint::make_router();
+            endpoint::serve_thread_main(router, http_listener).unwrap();
+        })
+        .unwrap();
+    threads.push(http_endpoint_thread);
+
     if conf.ttl.is_some() {
         let s3_conf = conf.clone();
         let s3_offload_thread = thread::Builder::new()

walkeeper/src/lib.rs

@@ -11,12 +11,23 @@ pub mod send_wal;
 pub mod timeline;
 pub mod wal_service;
 
+pub mod defaults {
+    use const_format::formatcp;
+
+    pub const DEFAULT_PG_LISTEN_PORT: u16 = 5454;
+    pub const DEFAULT_PG_LISTEN_ADDR: &str = formatcp!("127.0.0.1:{DEFAULT_PG_LISTEN_PORT}");
+
+    pub const DEFAULT_HTTP_LISTEN_PORT: u16 = 7676;
+    pub const DEFAULT_HTTP_LISTEN_ADDR: &str = formatcp!("127.0.0.1:{DEFAULT_HTTP_LISTEN_PORT}");
+}
+
 #[derive(Debug, Clone)]
 pub struct WalAcceptorConf {
     pub data_dir: PathBuf,
     pub daemonize: bool,
     pub no_sync: bool,
-    pub listen_addr: String,
+    pub listen_pg_addr: String,
+    pub listen_http_addr: String,
     pub pageserver_addr: Option<String>,
     // TODO (create issue) this is temporary, until protocol between PG<->SK<->PS rework
     pub pageserver_auth_token: Option<String>,

walkeeper/src/receive_wal.rs

@@ -42,7 +42,7 @@ fn request_callback(conf: WalAcceptorConf, timelineid: ZTimelineId, tenantid: ZT
     );
 
     // use Config parsing because SockAddr parsing doesnt allow to use host names instead of ip addresses
-    let me_connstr = format!("postgresql://no_user@{}/no_db", conf.listen_addr);
+    let me_connstr = format!("postgresql://no_user@{}/no_db", conf.listen_pg_addr);
     let me_conf: Config = me_connstr.parse().unwrap();
     let (host, port) = connection_host_port(&me_conf);
     let callme = format!(

walkeeper/src/safekeeper.rs

@@ -15,8 +15,11 @@ use std::cmp::min;
 use std::io;
 use std::io::Read;
 
+use lazy_static::lazy_static;
+
 use crate::replication::HotStandbyFeedback;
 use postgres_ffi::xlog_utils::MAX_SEND_SIZE;
+use zenith_metrics::{register_gauge_vec, Gauge, GaugeVec};
 use zenith_utils::bin_ser::LeSer;
 use zenith_utils::lsn::Lsn;
 use zenith_utils::pq_proto::SystemId;
@@ -279,6 +282,38 @@ pub trait Storage {
     fn write_wal(&mut self, server: &ServerInfo, startpos: Lsn, buf: &[u8]) -> Result<()>;
 }
 
+lazy_static! {
+    // The prometheus crate does not support u64 yet, i64 only (see `IntGauge`).
+    // i64 is faster than f64, so update to u64 when available.
+    static ref FLUSH_LSN_GAUGE: GaugeVec = register_gauge_vec!(
+        "safekeeper_flush_lsn",
+        "Current flush_lsn, grouped by timeline",
+        &["ztli"]
+    )
+    .expect("Failed to register safekeeper_flush_lsn int gauge vec");
+    static ref COMMIT_LSN_GAUGE: GaugeVec = register_gauge_vec!(
+        "safekeeper_commit_lsn",
+        "Current commit_lsn (not necessarily persisted to disk), grouped by timeline",
+        &["ztli"]
+    )
+    .expect("Failed to register safekeeper_commit_lsn int gauge vec");
+}
+
+struct SafeKeeperMetrics {
+    flush_lsn: Gauge,
+    commit_lsn: Gauge,
+}
+
+impl SafeKeeperMetrics {
+    fn new(ztli: ZTimelineId) -> SafeKeeperMetrics {
+        let ztli_str = format!("{}", ztli);
+        SafeKeeperMetrics {
+            flush_lsn: FLUSH_LSN_GAUGE.with_label_values(&[&ztli_str]),
+            commit_lsn: COMMIT_LSN_GAUGE.with_label_values(&[&ztli_str]),
+        }
+    }
+}
+
 /// SafeKeeper which consumes events (messages from compute) and provides
 /// replies.
 pub struct SafeKeeper<ST: Storage> {
@@ -286,6 +321,8 @@ pub struct SafeKeeper<ST: Storage> {
     /// Established by reading wal.
     pub flush_lsn: Lsn,
     pub tli: u32,
+    // Cached metrics so we don't have to recompute labels on each update.
+    metrics: Option<SafeKeeperMetrics>,
     /// not-yet-flushed pairs of same named fields in s.*
     pub commit_lsn: Lsn,
     pub truncate_lsn: Lsn,
@@ -304,6 +341,7 @@ where
         SafeKeeper {
             flush_lsn,
             tli,
+            metrics: None,
             commit_lsn: state.commit_lsn,
             truncate_lsn: state.truncate_lsn,
             storage,
@@ -355,6 +393,8 @@ where
         self.s.server.wal_seg_size = msg.wal_seg_size;
         self.storage.persist(&self.s, true)?;
 
+        self.metrics = Some(SafeKeeperMetrics::new(self.s.server.ztli));
+
         info!(
             "processed greeting from proposer {:?}, sending term {:?}",
             msg.proposer_id, self.s.acceptor_state.term
@@ -478,6 +518,11 @@ where
         }
         if last_rec_lsn > self.flush_lsn {
             self.flush_lsn = last_rec_lsn;
+            self.metrics
+                .as_ref()
+                .unwrap()
+                .flush_lsn
+                .set(u64::from(self.flush_lsn) as f64);
         }
 
         // Advance commit_lsn taking into account what we have locally. xxx this
@@ -495,6 +540,11 @@ where
             sync_control_file |=
                 commit_lsn >= msg.h.epoch_start_lsn && self.s.commit_lsn < msg.h.epoch_start_lsn;
             self.commit_lsn = commit_lsn;
+            self.metrics
+                .as_ref()
+                .unwrap()
+                .commit_lsn
+                .set(u64::from(self.commit_lsn) as f64);
         }
 
         self.truncate_lsn = msg.h.truncate_lsn;

walkeeper/src/wal_service.rs

@@ -13,9 +13,9 @@ use zenith_utils::postgres_backend::{AuthType, PostgresBackend};
 
 /// Accept incoming TCP connections and spawn them into a background thread.
 pub fn thread_main(conf: WalAcceptorConf) -> Result<()> {
-    info!("Starting wal acceptor on {}", conf.listen_addr);
-    let listener = TcpListener::bind(conf.listen_addr.clone()).map_err(|e| {
-        error!("failed to bind to address {}: {}", conf.listen_addr, e);
+    info!("Starting wal acceptor on {}", conf.listen_pg_addr);
+    let listener = TcpListener::bind(conf.listen_pg_addr.clone()).map_err(|e| {
+        error!("failed to bind to address {}: {}", conf.listen_pg_addr, e);
         e
     })?;
 

zenith_metrics/src/lib.rs

@@ -5,6 +5,8 @@
 use lazy_static::lazy_static;
 use once_cell::race::OnceBox;
 pub use prometheus::{exponential_buckets, linear_buckets};
+pub use prometheus::{register_gauge, Gauge};
+pub use prometheus::{register_gauge_vec, GaugeVec};
 pub use prometheus::{register_histogram, Histogram};
 pub use prometheus::{register_histogram_vec, HistogramVec};
 pub use prometheus::{register_int_counter, IntCounter};
@@ -21,7 +23,7 @@ pub use wrappers::{CountedReader, CountedWriter};
 /// Metrics gathering is a relatively simple and standalone operation, so
 /// it might be fine to do it this way to keep things simple.
 pub fn gather() -> Vec<prometheus::proto::MetricFamily> {
-    update_io_metrics();
+    update_rusage_metrics();
     prometheus::gather()
 }
 
@@ -31,16 +33,29 @@ static COMMON_METRICS_PREFIX: OnceBox<&str> = OnceBox::new();
 /// name like 'pageserver'. Should be executed exactly once in the beginning of
 /// any executable which uses common metrics.
 pub fn set_common_metrics_prefix(prefix: &'static str) {
-    COMMON_METRICS_PREFIX.set(prefix.into()).unwrap();
+    // Not unwrap() because metrics may be initialized after multiple threads have been started.
+    COMMON_METRICS_PREFIX
+        .set(prefix.into())
+        .unwrap_or_else(|_| {
+            eprintln!(
+                "set_common_metrics_prefix() was called second time with '{}', exiting",
+                prefix
+            );
+            std::process::exit(1);
+        });
 }
 
 /// Prepends a prefix to a common metric name so they are distinguished between
 /// different services, see https://github.com/zenithdb/zenith/pull/681
 /// A call to set_common_metrics_prefix() is necessary prior to calling this.
 pub fn new_common_metric_name(unprefixed_metric_name: &str) -> String {
+    // Not unwrap() because metrics may be initialized after multiple threads have been started.
     format!(
         "{}_{}",
-        COMMON_METRICS_PREFIX.get().unwrap(),
+        COMMON_METRICS_PREFIX.get().unwrap_or_else(|| {
+            eprintln!("set_common_metrics_prefix() was not called, but metrics are used, exiting");
+            std::process::exit(1);
+        }),
         unprefixed_metric_name
     )
 }
@@ -52,6 +67,11 @@ lazy_static! {
         &["io_operation"]
     )
     .expect("Failed to register disk i/o bytes int gauge vec");
+    static ref MAXRSS_KB: IntGauge = register_int_gauge!(
+        new_common_metric_name("maxrss_kb"),
+        "Memory usage (Maximum Resident Set Size)"
+    )
+    .expect("Failed to register maxrss_kb int gauge");
 }
 
 // Records I/O stats in a "cross-platform" way.
@@ -63,7 +83,7 @@ lazy_static! {
 // We know the size of the block, so we can determine the I/O bytes out of it.
 // The value might be not 100% exact, but should be fine for Prometheus metrics in this case.
 #[allow(clippy::unnecessary_cast)]
-fn update_io_metrics() {
+fn update_rusage_metrics() {
     let rusage_stats = get_rusage_stats();
 
     const BYTES_IN_BLOCK: i64 = 512;
@@ -73,6 +93,7 @@ fn update_io_metrics() {
     DISK_IO_BYTES
         .with_label_values(&["write"])
         .set(rusage_stats.ru_oublock * BYTES_IN_BLOCK);
+    MAXRSS_KB.set(rusage_stats.ru_maxrss);
 }
 
 fn get_rusage_stats() -> libc::rusage {

zenith_utils/src/lib.rs

@@ -8,6 +8,9 @@ pub mod lsn;
 /// SeqWait allows waiting for a future sequence number to arrive
 pub mod seqwait;
 
+/// append only ordered map implemented with a Vec
+pub mod vec_map;
+
 // Async version of SeqWait. Currently unused.
 // pub mod seqwait_async;
 

zenith_utils/src/vec_map.rs

@@ -0,0 +1,293 @@
+use std::{cmp::Ordering, ops::RangeBounds};
+
+use serde::{Deserialize, Serialize};
+
+/// Ordered map datastructure implemented in a Vec.
+/// Append only - can only add keys that are larger than the
+/// current max key.
+#[derive(Clone, Debug, Serialize, Deserialize)]
+pub struct VecMap<K, V>(Vec<(K, V)>);
+
+impl<K, V> Default for VecMap<K, V> {
+    fn default() -> Self {
+        VecMap(Default::default())
+    }
+}
+
+#[derive(Debug)]
+pub struct InvalidKey;
+
+impl<K: Ord, V> VecMap<K, V> {
+    pub fn is_empty(&self) -> bool {
+        self.0.is_empty()
+    }
+
+    pub fn as_slice(&self) -> &[(K, V)] {
+        self.0.as_slice()
+    }
+
+    /// This function may panic if given a range where the lower bound is
+    /// greater than the upper bound.
+    pub fn slice_range<R: RangeBounds<K>>(&self, range: R) -> &[(K, V)] {
+        use std::ops::Bound::*;
+
+        let binary_search = |k: &K| self.0.binary_search_by_key(&k, extract_key);
+
+        let start_idx = match range.start_bound() {
+            Unbounded => 0,
+            Included(k) => binary_search(k).unwrap_or_else(std::convert::identity),
+            Excluded(k) => match binary_search(k) {
+                Ok(idx) => idx + 1,
+                Err(idx) => idx,
+            },
+        };
+
+        let end_idx = match range.end_bound() {
+            Unbounded => self.0.len(),
+            Included(k) => match binary_search(k) {
+                Ok(idx) => idx + 1,
+                Err(idx) => idx,
+            },
+            Excluded(k) => binary_search(k).unwrap_or_else(std::convert::identity),
+        };
+
+        &self.0[start_idx..end_idx]
+    }
+
+    /// Add a key value pair to the map.
+    /// If [`key`] is less than or equal to the current maximum key
+    /// the pair will not be added and InvalidKey error will be returned.
+    pub fn append(&mut self, key: K, value: V) -> Result<(), InvalidKey> {
+        if let Some((last_key, _last_value)) = self.0.last() {
+            if &key <= last_key {
+                return Err(InvalidKey);
+            }
+        }
+
+        self.0.push((key, value));
+        Ok(())
+    }
+
+    /// Update the maximum key value pair or add a new key value pair to the map.
+    /// If [`key`] is less than the current maximum key no updates or additions
+    /// will occur and InvalidKey error will be returned.
+    pub fn append_or_update_last(&mut self, key: K, mut value: V) -> Result<Option<V>, InvalidKey> {
+        if let Some((last_key, last_value)) = self.0.last_mut() {
+            match key.cmp(last_key) {
+                Ordering::Less => return Err(InvalidKey),
+                Ordering::Equal => {
+                    std::mem::swap(last_value, &mut value);
+                    return Ok(Some(value));
+                }
+                Ordering::Greater => {}
+            }
+        }
+
+        self.0.push((key, value));
+        Ok(None)
+    }
+
+    /// Split the map into two.
+    ///
+    /// The left map contains everything before [`cutoff`] (exclusive).
+    /// Right map contains [`cutoff`] and everything after (inclusive).
+    pub fn split_at(&self, cutoff: &K) -> (Self, Self)
+    where
+        K: Clone,
+        V: Clone,
+    {
+        let split_idx = self
+            .0
+            .binary_search_by_key(&cutoff, extract_key)
+            .unwrap_or_else(std::convert::identity);
+
+        (
+            VecMap(self.0[..split_idx].to_vec()),
+            VecMap(self.0[split_idx..].to_vec()),
+        )
+    }
+
+    /// Move items from [`other`] to the end of [`self`], leaving [`other`] empty.
+    /// If any keys in [`other`] is less than or equal to any key in [`self`],
+    /// [`InvalidKey`] error will be returned and no mutation will occur.
+    pub fn extend(&mut self, other: &mut Self) -> Result<(), InvalidKey> {
+        let self_last_opt = self.0.last().map(extract_key);
+        let other_first_opt = other.0.last().map(extract_key);
+
+        if let (Some(self_last), Some(other_first)) = (self_last_opt, other_first_opt) {
+            if self_last >= other_first {
+                return Err(InvalidKey);
+            }
+        }
+
+        self.0.append(&mut other.0);
+
+        Ok(())
+    }
+}
+
+fn extract_key<K, V>(entry: &(K, V)) -> &K {
+    &entry.0
+}
+
+#[cfg(test)]
+mod tests {
+    use std::{collections::BTreeMap, ops::Bound};
+
+    use super::VecMap;
+
+    #[test]
+    fn unbounded_range() {
+        let mut vec = VecMap::default();
+        vec.append(0, ()).unwrap();
+
+        assert_eq!(vec.slice_range(0..0), &[]);
+    }
+
+    #[test]
+    #[should_panic]
+    fn invalid_ordering_range() {
+        let mut vec = VecMap::default();
+        vec.append(0, ()).unwrap();
+
+        #[allow(clippy::reversed_empty_ranges)]
+        vec.slice_range(1..0);
+    }
+
+    #[test]
+    fn range_tests() {
+        let mut vec = VecMap::default();
+        vec.append(0, ()).unwrap();
+        vec.append(2, ()).unwrap();
+        vec.append(4, ()).unwrap();
+
+        assert_eq!(vec.slice_range(0..0), &[]);
+        assert_eq!(vec.slice_range(0..1), &[(0, ())]);
+        assert_eq!(vec.slice_range(0..2), &[(0, ())]);
+        assert_eq!(vec.slice_range(0..3), &[(0, ()), (2, ())]);
+
+        assert_eq!(vec.slice_range(..0), &[]);
+        assert_eq!(vec.slice_range(..1), &[(0, ())]);
+
+        assert_eq!(vec.slice_range(..3), &[(0, ()), (2, ())]);
+        assert_eq!(vec.slice_range(..3), &[(0, ()), (2, ())]);
+
+        assert_eq!(vec.slice_range(0..=0), &[(0, ())]);
+        assert_eq!(vec.slice_range(0..=1), &[(0, ())]);
+        assert_eq!(vec.slice_range(0..=2), &[(0, ()), (2, ())]);
+        assert_eq!(vec.slice_range(0..=3), &[(0, ()), (2, ())]);
+
+        assert_eq!(vec.slice_range(..=0), &[(0, ())]);
+        assert_eq!(vec.slice_range(..=1), &[(0, ())]);
+        assert_eq!(vec.slice_range(..=2), &[(0, ()), (2, ())]);
+        assert_eq!(vec.slice_range(..=3), &[(0, ()), (2, ())]);
+    }
+
+    struct BoundIter {
+        min: i32,
+        max: i32,
+
+        next: Option<Bound<i32>>,
+    }
+
+    impl BoundIter {
+        fn new(min: i32, max: i32) -> Self {
+            Self {
+                min,
+                max,
+
+                next: Some(Bound::Unbounded),
+            }
+        }
+    }
+
+    impl Iterator for BoundIter {
+        type Item = Bound<i32>;
+
+        fn next(&mut self) -> Option<Self::Item> {
+            let cur = self.next?;
+
+            self.next = match &cur {
+                Bound::Unbounded => Some(Bound::Included(self.min)),
+                Bound::Included(x) => {
+                    if *x >= self.max {
+                        Some(Bound::Excluded(self.min))
+                    } else {
+                        Some(Bound::Included(x + 1))
+                    }
+                }
+                Bound::Excluded(x) => {
+                    if *x >= self.max {
+                        None
+                    } else {
+                        Some(Bound::Excluded(x + 1))
+                    }
+                }
+            };
+
+            Some(cur)
+        }
+    }
+
+    #[test]
+    fn range_exhaustive() {
+        let map: BTreeMap<i32, ()> = (1..=7).step_by(2).map(|x| (x, ())).collect();
+        let mut vec = VecMap::default();
+        for &key in map.keys() {
+            vec.append(key, ()).unwrap();
+        }
+
+        const RANGE_MIN: i32 = 0;
+        const RANGE_MAX: i32 = 8;
+        for lower_bound in BoundIter::new(RANGE_MIN, RANGE_MAX) {
+            let ub_min = match lower_bound {
+                Bound::Unbounded => RANGE_MIN,
+                Bound::Included(x) => x,
+                Bound::Excluded(x) => x + 1,
+            };
+            for upper_bound in BoundIter::new(ub_min, RANGE_MAX) {
+                let map_range: Vec<(i32, ())> = map
+                    .range((lower_bound, upper_bound))
+                    .map(|(&x, _)| (x, ()))
+                    .collect();
+                let vec_slice = vec.slice_range((lower_bound, upper_bound));
+
+                assert_eq!(map_range, vec_slice);
+            }
+        }
+    }
+
+    #[test]
+    fn extend() {
+        let mut left = VecMap::default();
+        left.append(0, ()).unwrap();
+        assert_eq!(left.as_slice(), &[(0, ())]);
+
+        let mut empty = VecMap::default();
+        left.extend(&mut empty).unwrap();
+        assert_eq!(left.as_slice(), &[(0, ())]);
+        assert_eq!(empty.as_slice(), &[]);
+
+        let mut right = VecMap::default();
+        right.append(1, ()).unwrap();
+
+        left.extend(&mut right).unwrap();
+
+        assert_eq!(left.as_slice(), &[(0, ()), (1, ())]);
+        assert_eq!(right.as_slice(), &[]);
+
+        let mut zero_map = VecMap::default();
+        zero_map.append(0, ()).unwrap();
+
+        left.extend(&mut zero_map).unwrap_err();
+        assert_eq!(left.as_slice(), &[(0, ()), (1, ())]);
+        assert_eq!(zero_map.as_slice(), &[(0, ())]);
+
+        let mut one_map = VecMap::default();
+        one_map.append(1, ()).unwrap();
+
+        left.extend(&mut one_map).unwrap_err();
+        assert_eq!(left.as_slice(), &[(0, ()), (1, ())]);
+        assert_eq!(one_map.as_slice(), &[(1, ())]);
+    }
+}