safekeeper: use protobuf for sending compressed records to pageserver (…

…#9821)

## Problem

#9746 lifted decoding and
interpretation of WAL to the safekeeper.
This reduced the ingested amount on the pageservers by around 10x for a
tenant with 8 shards, but doubled
the ingested amount for single sharded tenants.

Also, #9746 uses bincode which
doesn't support schema evolution.
Technically the schema can be evolved, but it's very cumbersome.

## Summary of changes

This patch set addresses both problems by adding protobuf support for
the interpreted wal records and adding compression support. Compressed
protobuf reduced the ingested amount by 100x on the 32 shards
`test_sharded_ingest` case (compared to non-interpreted proto). For the
1 shard case the reduction is 5x.

Sister change to `rust-postgres` is
[here](neondatabase/rust-postgres#33).

## Links

Related: #9336
Epic: #9329

libs/pageserver_api/src/key.rs

@@ -229,6 +229,18 @@ impl Key {
     }
 }
 
+impl CompactKey {
+    pub fn raw(&self) -> i128 {
+        self.0
+    }
+}
+
+impl From<i128> for CompactKey {
+    fn from(value: i128) -> Self {
+        Self(value)
+    }
+}
+
 impl fmt::Display for Key {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
         write!(

libs/pq_proto/src/lib.rs

@@ -688,9 +688,6 @@ pub struct InterpretedWalRecordsBody<'a> {
     pub streaming_lsn: u64,
     /// Current end of WAL on the server
     pub commit_lsn: u64,
-    /// Start LSN of the next record in PG WAL.
-    /// Is 0 if the portion of PG WAL did not contain any records.
-    pub next_record_lsn: u64,
     pub data: &'a [u8],
 }
 
@@ -1028,7 +1025,6 @@ impl BeMessage<'_> {
                                       // dependency
                     buf.put_u64(rec.streaming_lsn);
                     buf.put_u64(rec.commit_lsn);
-                    buf.put_u64(rec.next_record_lsn);
                     buf.put_slice(rec.data);
                 });
             }

libs/utils/src/postgres_client.rs

@@ -7,40 +7,31 @@ use postgres_connection::{parse_host_port, PgConnectionConfig};
 
 use crate::id::TenantTimelineId;
 
-/// Postgres client protocol types
-#[derive(
-    Copy,
-    Clone,
-    PartialEq,
-    Eq,
-    strum_macros::EnumString,
-    strum_macros::Display,
-    serde_with::DeserializeFromStr,
-    serde_with::SerializeDisplay,
-    Debug,
-)]
-#[strum(serialize_all = "kebab-case")]
-#[repr(u8)]
+#[derive(Copy, Clone, PartialEq, Eq, Debug, serde::Serialize, serde::Deserialize)]
+#[serde(rename_all = "kebab-case")]
+pub enum InterpretedFormat {
+    Bincode,
+    Protobuf,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+#[serde(rename_all = "kebab-case")]
+pub enum Compression {
+    Zstd { level: i8 },
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+#[serde(tag = "type", content = "args")]
+#[serde(rename_all = "kebab-case")]
 pub enum PostgresClientProtocol {
     /// Usual Postgres replication protocol
     Vanilla,
     /// Custom shard-aware protocol that replicates interpreted records.
     /// Used to send wal from safekeeper to pageserver.
-    Interpreted,
-}
-
-impl TryFrom<u8> for PostgresClientProtocol {
-    type Error = u8;
-
-    fn try_from(value: u8) -> Result<Self, Self::Error> {
-        Ok(match value {
-            v if v == (PostgresClientProtocol::Vanilla as u8) => PostgresClientProtocol::Vanilla,
-            v if v == (PostgresClientProtocol::Interpreted as u8) => {
-                PostgresClientProtocol::Interpreted
-            }
-            x => return Err(x),
-        })
-    }
+    Interpreted {
+        format: InterpretedFormat,
+        compression: Option<Compression>,
+    },
 }
 
 pub struct ConnectionConfigArgs<'a> {
@@ -63,7 +54,10 @@ impl<'a> ConnectionConfigArgs<'a> {
             "-c".to_owned(),
             format!("timeline_id={}", self.ttid.timeline_id),
             format!("tenant_id={}", self.ttid.tenant_id),
-            format!("protocol={}", self.protocol as u8),
+            format!(
+                "protocol={}",
+                serde_json::to_string(&self.protocol).unwrap()
+            ),
         ];
 
         if self.shard_number.is_some() {

libs/wal_decoder/Cargo.toml

@@ -8,11 +8,19 @@ license.workspace = true
 testing = ["pageserver_api/testing"]
 
 [dependencies]
+async-compression.workspace = true
 anyhow.workspace = true
 bytes.workspace = true
 pageserver_api.workspace = true
+prost.workspace = true
 postgres_ffi.workspace = true
 serde.workspace = true
+thiserror.workspace = true
+tokio = { workspace = true, features = ["io-util"] }
+tonic.workspace = true
 tracing.workspace = true
 utils.workspace = true
 workspace_hack = { version = "0.1", path = "../../workspace_hack" }
+
+[build-dependencies]
+tonic-build.workspace = true

libs/wal_decoder/build.rs

@@ -0,0 +1,11 @@
+fn main() -> Result<(), Box<dyn std::error::Error>> {
+    // Generate rust code from .proto protobuf.
+    //
+    // Note: we previously tried to use deterministic location at proto/ for
+    // easy location, but apparently interference with cachepot sometimes fails
+    // the build then. Anyway, per cargo docs build script shouldn't output to
+    // anywhere but $OUT_DIR.
+    tonic_build::compile_protos("proto/interpreted_wal.proto")
+        .unwrap_or_else(|e| panic!("failed to compile protos {:?}", e));
+    Ok(())
+}

libs/wal_decoder/proto/interpreted_wal.proto

@@ -0,0 +1,43 @@
+syntax = "proto3";
+
+package interpreted_wal;
+
+message InterpretedWalRecords {
+  repeated InterpretedWalRecord records = 1;
+  optional uint64 next_record_lsn = 2;
+}
+
+message InterpretedWalRecord {
+  optional bytes metadata_record = 1;
+  SerializedValueBatch batch = 2;
+  uint64 next_record_lsn = 3;
+  bool flush_uncommitted = 4;
+  uint32 xid = 5;
+}
+
+message SerializedValueBatch {
+  bytes raw = 1;
+  repeated ValueMeta metadata = 2;
+  uint64 max_lsn = 3;
+  uint64 len = 4;
+}
+
+enum ValueMetaType {
+  Serialized = 0;
+  Observed = 1;
+}
+
+message ValueMeta {
+  ValueMetaType type = 1;
+  CompactKey key = 2;
+  uint64 lsn = 3;
+  optional uint64 batch_offset = 4;
+  optional uint64 len = 5;
+  optional bool will_init = 6;
+}
+
+message CompactKey {
+  int64 high = 1;
+  int64 low = 2;
+}
+

libs/wal_decoder/src/lib.rs

@@ -1,3 +1,4 @@
 pub mod decoder;
 pub mod models;
 pub mod serialized_batch;
+pub mod wire_format;

libs/wal_decoder/src/models.rs

@@ -37,12 +37,32 @@ use utils::lsn::Lsn;
 
 use crate::serialized_batch::SerializedValueBatch;
 
+// Code generated by protobuf.
+pub mod proto {
+    // Tonic does derives as `#[derive(Clone, PartialEq, ::prost::Message)]`
+    // we don't use these types for anything but broker data transmission,
+    // so it's ok to ignore this one.
+    #![allow(clippy::derive_partial_eq_without_eq)]
+    // The generated ValueMeta has a `len` method generate for its `len` field.
+    #![allow(clippy::len_without_is_empty)]
+    tonic::include_proto!("interpreted_wal");
+}
+
 #[derive(Serialize, Deserialize)]
 pub enum FlushUncommittedRecords {
     Yes,
     No,
 }
 
+/// A batch of interpreted WAL records
+#[derive(Serialize, Deserialize)]
+pub struct InterpretedWalRecords {
+    pub records: Vec<InterpretedWalRecord>,
+    // Start LSN of the next record after the batch.
+    // Note that said record may not belong to the current shard.
+    pub next_record_lsn: Option<Lsn>,
+}
+
 /// An interpreted Postgres WAL record, ready to be handled by the pageserver
 #[derive(Serialize, Deserialize)]
 pub struct InterpretedWalRecord {