Proxy release 2025-07-03 10:07 UTC

## Problem

`keep_connection` does not exit, so it was never setting
`credentials_refreshed`.

## Summary of changes

Set `credentials_refreshed` to true when we first establish a
connection, and after we re-authenticate the connection.

.gitignore

@@ -6,7 +6,6 @@
 /tmp_check_cli
 __pycache__/
 test_output/
-neon_previous/
 .vscode
 .idea
 *.swp

Cargo.lock

@@ -1305,7 +1305,6 @@ dependencies = [
  "fail",
  "flate2",
  "futures",
- "hostname-validator",
  "http 1.1.0",
  "indexmap 2.9.0",
  "itertools 0.10.5",
@@ -2772,12 +2771,6 @@ dependencies = [
  "windows",
 ]
 
-[[package]]
-name = "hostname-validator"
-version = "1.1.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f558a64ac9af88b5ba400d99b579451af0d39c6d360980045b91aac966d705e2"
-
 [[package]]
 name = "http"
 version = "0.2.9"

compute/compute-node.Dockerfile

@@ -1572,7 +1572,6 @@ RUN make -j $(getconf _NPROCESSORS_ONLN) && \
 FROM build-deps AS pgaudit-src
 ARG PG_VERSION
 WORKDIR /ext-src
-COPY "compute/patches/pgaudit-parallel_workers-${PG_VERSION}.patch" .
 RUN case "${PG_VERSION}" in \
     "v14") \
     export PGAUDIT_VERSION=1.6.3 \
@@ -1595,8 +1594,7 @@ RUN case "${PG_VERSION}" in \
     esac && \
     wget https://github.com/pgaudit/pgaudit/archive/refs/tags/${PGAUDIT_VERSION}.tar.gz -O pgaudit.tar.gz && \
     echo "${PGAUDIT_CHECKSUM} pgaudit.tar.gz" | sha256sum --check && \
-    mkdir pgaudit-src && cd pgaudit-src && tar xzf ../pgaudit.tar.gz --strip-components=1 -C . && \
-    patch -p1 < "/ext-src/pgaudit-parallel_workers-${PG_VERSION}.patch"
+    mkdir pgaudit-src && cd pgaudit-src && tar xzf ../pgaudit.tar.gz --strip-components=1 -C .
 
 FROM pg-build AS pgaudit-build
 COPY --from=pgaudit-src /ext-src/ /ext-src/
@@ -1985,7 +1983,7 @@ RUN apt update && \
         locales \
         lsof \
         procps \
-        rsyslog-gnutls \
+        rsyslog \
         screen \
         tcpdump \
         $VERSION_INSTALLS && \

compute/etc/neon_collector.jsonnet

@@ -8,8 +8,6 @@
     import 'sql_exporter/compute_logical_snapshot_files.libsonnet',
     import 'sql_exporter/compute_logical_snapshots_bytes.libsonnet',
     import 'sql_exporter/compute_max_connections.libsonnet',
-    import 'sql_exporter/compute_pg_oldest_frozen_xid_age.libsonnet',
-    import 'sql_exporter/compute_pg_oldest_mxid_age.libsonnet',
     import 'sql_exporter/compute_receive_lsn.libsonnet',
     import 'sql_exporter/compute_subscriptions_count.libsonnet',
     import 'sql_exporter/connection_counts.libsonnet',

compute/etc/sql_exporter/compute_pg_oldest_frozen_xid_age.libsonnet

@@ -1,13 +0,0 @@
-{
-  metric_name: 'compute_pg_oldest_frozen_xid_age',
-  type: 'gauge',
-  help: 'Age of oldest XIDs that have not been frozen by VACUUM. An indicator of how long it has been since VACUUM last ran.',
-  key_labels: [
-    'database_name',
-  ],
-  value_label: 'metric',
-  values: [
-    'frozen_xid_age',
-  ],
-  query: importstr 'sql_exporter/compute_pg_oldest_frozen_xid_age.sql',
-}

compute/etc/sql_exporter/compute_pg_oldest_frozen_xid_age.sql

@@ -1,4 +0,0 @@
-SELECT datname database_name,
-  age(datfrozenxid) frozen_xid_age
-FROM pg_database
-ORDER BY frozen_xid_age DESC LIMIT 10;

compute/etc/sql_exporter/compute_pg_oldest_mxid_age.libsonnet

@@ -1,13 +0,0 @@
-{
-  metric_name: 'compute_pg_oldest_mxid_age',
-  type: 'gauge',
-  help: 'Age of oldest MXIDs that have not been replaced by VACUUM. An indicator of how long it has been since VACUUM last ran.',
-  key_labels: [
-    'database_name',
-  ],
-  value_label: 'metric',
-  values: [
-    'min_mxid_age',
-  ],
-  query: importstr 'sql_exporter/compute_pg_oldest_mxid_age.sql',
-}

compute/etc/sql_exporter/compute_pg_oldest_mxid_age.sql

@@ -1,4 +0,0 @@
-SELECT datname database_name,
-  mxid_age(datminmxid) min_mxid_age
-FROM pg_database
-ORDER BY min_mxid_age DESC LIMIT 10;

compute/patches/anon_v2.patch

@@ -1,8 +1,8 @@
 diff --git a/sql/anon.sql b/sql/anon.sql
-index 0cdc769..b450327 100644
+index 0cdc769..f6cc950 100644
 --- a/sql/anon.sql
 +++ b/sql/anon.sql
-@@ -1141,3 +1141,15 @@ $$
+@@ -1141,3 +1141,8 @@ $$
  -- TODO : https://en.wikipedia.org/wiki/L-diversity
  
  -- TODO : https://en.wikipedia.org/wiki/T-closeness
@@ -11,13 +11,6 @@ index 0cdc769..b450327 100644
 +
 +GRANT ALL ON SCHEMA anon to neon_superuser;
 +GRANT ALL ON ALL TABLES IN SCHEMA anon TO neon_superuser;
-+
-+DO $$
-+BEGIN
-+    IF current_setting('server_version_num')::int >= 150000 THEN
-+        GRANT SET ON PARAMETER anon.transparent_dynamic_masking TO neon_superuser;
-+    END IF;
-+END $$;
 diff --git a/sql/init.sql b/sql/init.sql
 index 7da6553..9b6164b 100644
 --- a/sql/init.sql

compute/patches/pgaudit-parallel_workers-v14.patch

@@ -1,143 +0,0 @@
-commit 7220bb3a3f23fa27207d77562dcc286f9a123313
-Author: Tristan Partin <tristan.partin@databricks.com>
-Date:   2025-06-23 02:09:31 +0000
-
-    Disable logging in parallel workers
-    
-    When a query uses parallel workers, pgaudit will log the same query for
-    every parallel worker. This is undesireable since it can result in log
-    amplification for queries that use parallel workers.
-    
-    Signed-off-by: Tristan Partin <tristan.partin@databricks.com>
-
-diff --git a/expected/pgaudit.out b/expected/pgaudit.out
-index baa8011..a601375 100644
---- a/expected/pgaudit.out
-+++ b/expected/pgaudit.out
-@@ -2563,6 +2563,37 @@ COMMIT;
- NOTICE:  AUDIT: SESSION,12,4,MISC,COMMIT,,,COMMIT;,<not logged>
- DROP TABLE part_test;
- NOTICE:  AUDIT: SESSION,13,1,DDL,DROP TABLE,,,DROP TABLE part_test;,<not logged>
-+--
-+-- Test logging in parallel workers
-+SET pgaudit.log = 'read';
-+SET pgaudit.log_client = on;
-+SET pgaudit.log_level = 'notice';
-+-- Force parallel execution for testing
-+SET max_parallel_workers_per_gather = 2;
-+SET parallel_tuple_cost = 0;
-+SET parallel_setup_cost = 0;
-+SET min_parallel_table_scan_size = 0;
-+SET min_parallel_index_scan_size = 0;
-+-- Create table with enough data to trigger parallel execution
-+CREATE TABLE parallel_test (id int, data text);
-+INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
-+SELECT count(*) FROM parallel_test;
-+NOTICE:  AUDIT: SESSION,14,1,READ,SELECT,,,SELECT count(*) FROM parallel_test;,<not logged>
-+ count 
-+-------
-+  1000
-+(1 row)
-+
-+-- Cleanup parallel test
-+DROP TABLE parallel_test;
-+RESET max_parallel_workers_per_gather;
-+RESET parallel_tuple_cost;
-+RESET parallel_setup_cost;
-+RESET min_parallel_table_scan_size;
-+RESET min_parallel_index_scan_size;
-+RESET pgaudit.log;
-+RESET pgaudit.log_client;
-+RESET pgaudit.log_level;
- -- Cleanup
- -- Set client_min_messages up to warning to avoid noise
- SET client_min_messages = 'warning';
-diff --git a/pgaudit.c b/pgaudit.c
-index 5e6fd38..ac9ded2 100644
---- a/pgaudit.c
-+++ b/pgaudit.c
-@@ -11,6 +11,7 @@
- #include "postgres.h"
- 
- #include "access/htup_details.h"
-+#include "access/parallel.h"
- #include "access/sysattr.h"
- #include "access/xact.h"
- #include "access/relation.h"
-@@ -1303,7 +1304,7 @@ pgaudit_ExecutorStart_hook(QueryDesc *queryDesc, int eflags)
- {
-     AuditEventStackItem *stackItem = NULL;
- 
--    if (!internalStatement)
-+    if (!internalStatement && !IsParallelWorker())
-     {
-         /* Push the audit even onto the stack */
-         stackItem = stack_push();
-@@ -1384,7 +1385,7 @@ pgaudit_ExecutorCheckPerms_hook(List *rangeTabls, bool abort)
- 
-     /* Log DML if the audit role is valid or session logging is enabled */
-     if ((auditOid != InvalidOid || auditLogBitmap != 0) &&
--        !IsAbortedTransactionBlockState())
-+        !IsAbortedTransactionBlockState() && !IsParallelWorker())
-     {
-         /* If auditLogRows is on, wait for rows processed to be set */
-         if (auditLogRows && auditEventStack != NULL)
-@@ -1438,7 +1439,7 @@ pgaudit_ExecutorRun_hook(QueryDesc *queryDesc, ScanDirection direction, uint64 c
-     else
-         standard_ExecutorRun(queryDesc, direction, count, execute_once);
- 
--    if (auditLogRows && !internalStatement)
-+    if (auditLogRows && !internalStatement && !IsParallelWorker())
-     {
-         /* Find an item from the stack by the query memory context */
-         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
-@@ -1458,7 +1459,7 @@ pgaudit_ExecutorEnd_hook(QueryDesc *queryDesc)
-     AuditEventStackItem *stackItem = NULL;
-     AuditEventStackItem *auditEventStackFull = NULL;
- 
--    if (auditLogRows && !internalStatement)
-+    if (auditLogRows && !internalStatement && !IsParallelWorker())
-     {
-         /* Find an item from the stack by the query memory context */
-         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
-diff --git a/sql/pgaudit.sql b/sql/pgaudit.sql
-index cc1374a..1870a60 100644
---- a/sql/pgaudit.sql
-+++ b/sql/pgaudit.sql
-@@ -1612,6 +1612,36 @@ COMMIT;
- 
- DROP TABLE part_test;
- 
-+--
-+-- Test logging in parallel workers
-+SET pgaudit.log = 'read';
-+SET pgaudit.log_client = on;
-+SET pgaudit.log_level = 'notice';
-+
-+-- Force parallel execution for testing
-+SET max_parallel_workers_per_gather = 2;
-+SET parallel_tuple_cost = 0;
-+SET parallel_setup_cost = 0;
-+SET min_parallel_table_scan_size = 0;
-+SET min_parallel_index_scan_size = 0;
-+
-+-- Create table with enough data to trigger parallel execution
-+CREATE TABLE parallel_test (id int, data text);
-+INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
-+
-+SELECT count(*) FROM parallel_test;
-+
-+-- Cleanup parallel test
-+DROP TABLE parallel_test;
-+RESET max_parallel_workers_per_gather;
-+RESET parallel_tuple_cost;
-+RESET parallel_setup_cost;
-+RESET min_parallel_table_scan_size;
-+RESET min_parallel_index_scan_size;
-+RESET pgaudit.log;
-+RESET pgaudit.log_client;
-+RESET pgaudit.log_level;
-+
- -- Cleanup
- -- Set client_min_messages up to warning to avoid noise
- SET client_min_messages = 'warning';

compute/patches/pgaudit-parallel_workers-v15.patch

@@ -1,143 +0,0 @@
-commit 29dc2847f6255541992f18faf8a815dfab79631a
-Author: Tristan Partin <tristan.partin@databricks.com>
-Date:   2025-06-23 02:09:31 +0000
-
-    Disable logging in parallel workers
-    
-    When a query uses parallel workers, pgaudit will log the same query for
-    every parallel worker. This is undesireable since it can result in log
-    amplification for queries that use parallel workers.
-    
-    Signed-off-by: Tristan Partin <tristan.partin@databricks.com>
-
-diff --git a/expected/pgaudit.out b/expected/pgaudit.out
-index b22560b..73f0327 100644
---- a/expected/pgaudit.out
-+++ b/expected/pgaudit.out
-@@ -2563,6 +2563,37 @@ COMMIT;
- NOTICE:  AUDIT: SESSION,12,4,MISC,COMMIT,,,COMMIT;,<not logged>
- DROP TABLE part_test;
- NOTICE:  AUDIT: SESSION,13,1,DDL,DROP TABLE,,,DROP TABLE part_test;,<not logged>
-+--
-+-- Test logging in parallel workers
-+SET pgaudit.log = 'read';
-+SET pgaudit.log_client = on;
-+SET pgaudit.log_level = 'notice';
-+-- Force parallel execution for testing
-+SET max_parallel_workers_per_gather = 2;
-+SET parallel_tuple_cost = 0;
-+SET parallel_setup_cost = 0;
-+SET min_parallel_table_scan_size = 0;
-+SET min_parallel_index_scan_size = 0;
-+-- Create table with enough data to trigger parallel execution
-+CREATE TABLE parallel_test (id int, data text);
-+INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
-+SELECT count(*) FROM parallel_test;
-+NOTICE:  AUDIT: SESSION,14,1,READ,SELECT,,,SELECT count(*) FROM parallel_test;,<not logged>
-+ count 
-+-------
-+  1000
-+(1 row)
-+
-+-- Cleanup parallel test
-+DROP TABLE parallel_test;
-+RESET max_parallel_workers_per_gather;
-+RESET parallel_tuple_cost;
-+RESET parallel_setup_cost;
-+RESET min_parallel_table_scan_size;
-+RESET min_parallel_index_scan_size;
-+RESET pgaudit.log;
-+RESET pgaudit.log_client;
-+RESET pgaudit.log_level;
- -- Cleanup
- -- Set client_min_messages up to warning to avoid noise
- SET client_min_messages = 'warning';
-diff --git a/pgaudit.c b/pgaudit.c
-index 5e6fd38..ac9ded2 100644
---- a/pgaudit.c
-+++ b/pgaudit.c
-@@ -11,6 +11,7 @@
- #include "postgres.h"
- 
- #include "access/htup_details.h"
-+#include "access/parallel.h"
- #include "access/sysattr.h"
- #include "access/xact.h"
- #include "access/relation.h"
-@@ -1303,7 +1304,7 @@ pgaudit_ExecutorStart_hook(QueryDesc *queryDesc, int eflags)
- {
-     AuditEventStackItem *stackItem = NULL;
- 
--    if (!internalStatement)
-+    if (!internalStatement && !IsParallelWorker())
-     {
-         /* Push the audit even onto the stack */
-         stackItem = stack_push();
-@@ -1384,7 +1385,7 @@ pgaudit_ExecutorCheckPerms_hook(List *rangeTabls, bool abort)
- 
-     /* Log DML if the audit role is valid or session logging is enabled */
-     if ((auditOid != InvalidOid || auditLogBitmap != 0) &&
--        !IsAbortedTransactionBlockState())
-+        !IsAbortedTransactionBlockState() && !IsParallelWorker())
-     {
-         /* If auditLogRows is on, wait for rows processed to be set */
-         if (auditLogRows && auditEventStack != NULL)
-@@ -1438,7 +1439,7 @@ pgaudit_ExecutorRun_hook(QueryDesc *queryDesc, ScanDirection direction, uint64 c
-     else
-         standard_ExecutorRun(queryDesc, direction, count, execute_once);
- 
--    if (auditLogRows && !internalStatement)
-+    if (auditLogRows && !internalStatement && !IsParallelWorker())
-     {
-         /* Find an item from the stack by the query memory context */
-         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
-@@ -1458,7 +1459,7 @@ pgaudit_ExecutorEnd_hook(QueryDesc *queryDesc)
-     AuditEventStackItem *stackItem = NULL;
-     AuditEventStackItem *auditEventStackFull = NULL;
- 
--    if (auditLogRows && !internalStatement)
-+    if (auditLogRows && !internalStatement && !IsParallelWorker())
-     {
-         /* Find an item from the stack by the query memory context */
-         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
-diff --git a/sql/pgaudit.sql b/sql/pgaudit.sql
-index 8052426..7f0667b 100644
---- a/sql/pgaudit.sql
-+++ b/sql/pgaudit.sql
-@@ -1612,6 +1612,36 @@ COMMIT;
- 
- DROP TABLE part_test;
- 
-+--
-+-- Test logging in parallel workers
-+SET pgaudit.log = 'read';
-+SET pgaudit.log_client = on;
-+SET pgaudit.log_level = 'notice';
-+
-+-- Force parallel execution for testing
-+SET max_parallel_workers_per_gather = 2;
-+SET parallel_tuple_cost = 0;
-+SET parallel_setup_cost = 0;
-+SET min_parallel_table_scan_size = 0;
-+SET min_parallel_index_scan_size = 0;
-+
-+-- Create table with enough data to trigger parallel execution
-+CREATE TABLE parallel_test (id int, data text);
-+INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
-+
-+SELECT count(*) FROM parallel_test;
-+
-+-- Cleanup parallel test
-+DROP TABLE parallel_test;
-+RESET max_parallel_workers_per_gather;
-+RESET parallel_tuple_cost;
-+RESET parallel_setup_cost;
-+RESET min_parallel_table_scan_size;
-+RESET min_parallel_index_scan_size;
-+RESET pgaudit.log;
-+RESET pgaudit.log_client;
-+RESET pgaudit.log_level;
-+
- -- Cleanup
- -- Set client_min_messages up to warning to avoid noise
- SET client_min_messages = 'warning';

compute/patches/pgaudit-parallel_workers-v16.patch

@@ -1,143 +0,0 @@
-commit cc708dde7ef2af2a8120d757102d2e34c0463a0f
-Author: Tristan Partin <tristan.partin@databricks.com>
-Date:   2025-06-23 02:09:31 +0000
-
-    Disable logging in parallel workers
-    
-    When a query uses parallel workers, pgaudit will log the same query for
-    every parallel worker. This is undesireable since it can result in log
-    amplification for queries that use parallel workers.
-    
-    Signed-off-by: Tristan Partin <tristan.partin@databricks.com>
-
-diff --git a/expected/pgaudit.out b/expected/pgaudit.out
-index 8772054..9b66ac6 100644
---- a/expected/pgaudit.out
-+++ b/expected/pgaudit.out
-@@ -2556,6 +2556,37 @@ DROP SERVER fdw_server;
- NOTICE:  AUDIT: SESSION,11,1,DDL,DROP SERVER,,,DROP SERVER fdw_server;,<not logged>
- DROP EXTENSION postgres_fdw;
- NOTICE:  AUDIT: SESSION,12,1,DDL,DROP EXTENSION,,,DROP EXTENSION postgres_fdw;,<not logged>
-+--
-+-- Test logging in parallel workers
-+SET pgaudit.log = 'read';
-+SET pgaudit.log_client = on;
-+SET pgaudit.log_level = 'notice';
-+-- Force parallel execution for testing
-+SET max_parallel_workers_per_gather = 2;
-+SET parallel_tuple_cost = 0;
-+SET parallel_setup_cost = 0;
-+SET min_parallel_table_scan_size = 0;
-+SET min_parallel_index_scan_size = 0;
-+-- Create table with enough data to trigger parallel execution
-+CREATE TABLE parallel_test (id int, data text);
-+INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
-+SELECT count(*) FROM parallel_test;
-+NOTICE:  AUDIT: SESSION,13,1,READ,SELECT,,,SELECT count(*) FROM parallel_test;,<not logged>
-+ count 
-+-------
-+  1000
-+(1 row)
-+
-+-- Cleanup parallel test
-+DROP TABLE parallel_test;
-+RESET max_parallel_workers_per_gather;
-+RESET parallel_tuple_cost;
-+RESET parallel_setup_cost;
-+RESET min_parallel_table_scan_size;
-+RESET min_parallel_index_scan_size;
-+RESET pgaudit.log;
-+RESET pgaudit.log_client;
-+RESET pgaudit.log_level;
- -- Cleanup
- -- Set client_min_messages up to warning to avoid noise
- SET client_min_messages = 'warning';
-diff --git a/pgaudit.c b/pgaudit.c
-index 004d1f9..f061164 100644
---- a/pgaudit.c
-+++ b/pgaudit.c
-@@ -11,6 +11,7 @@
- #include "postgres.h"
- 
- #include "access/htup_details.h"
-+#include "access/parallel.h"
- #include "access/sysattr.h"
- #include "access/xact.h"
- #include "access/relation.h"
-@@ -1339,7 +1340,7 @@ pgaudit_ExecutorStart_hook(QueryDesc *queryDesc, int eflags)
- {
-     AuditEventStackItem *stackItem = NULL;
- 
--    if (!internalStatement)
-+    if (!internalStatement && !IsParallelWorker())
-     {
-         /* Push the audit even onto the stack */
-         stackItem = stack_push();
-@@ -1420,7 +1421,7 @@ pgaudit_ExecutorCheckPerms_hook(List *rangeTabls, List *permInfos, bool abort)
- 
-     /* Log DML if the audit role is valid or session logging is enabled */
-     if ((auditOid != InvalidOid || auditLogBitmap != 0) &&
--        !IsAbortedTransactionBlockState())
-+        !IsAbortedTransactionBlockState() && !IsParallelWorker())
-     {
-         /* If auditLogRows is on, wait for rows processed to be set */
-         if (auditLogRows && auditEventStack != NULL)
-@@ -1475,7 +1476,7 @@ pgaudit_ExecutorRun_hook(QueryDesc *queryDesc, ScanDirection direction, uint64 c
-     else
-         standard_ExecutorRun(queryDesc, direction, count, execute_once);
- 
--    if (auditLogRows && !internalStatement)
-+    if (auditLogRows && !internalStatement && !IsParallelWorker())
-     {
-         /* Find an item from the stack by the query memory context */
-         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
-@@ -1495,7 +1496,7 @@ pgaudit_ExecutorEnd_hook(QueryDesc *queryDesc)
-     AuditEventStackItem *stackItem = NULL;
-     AuditEventStackItem *auditEventStackFull = NULL;
- 
--    if (auditLogRows && !internalStatement)
-+    if (auditLogRows && !internalStatement && !IsParallelWorker())
-     {
-         /* Find an item from the stack by the query memory context */
-         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
-diff --git a/sql/pgaudit.sql b/sql/pgaudit.sql
-index 6aae88b..de6d7fd 100644
---- a/sql/pgaudit.sql
-+++ b/sql/pgaudit.sql
-@@ -1631,6 +1631,36 @@ DROP USER MAPPING FOR regress_user1 SERVER fdw_server;
- DROP SERVER fdw_server;
- DROP EXTENSION postgres_fdw;
- 
-+--
-+-- Test logging in parallel workers
-+SET pgaudit.log = 'read';
-+SET pgaudit.log_client = on;
-+SET pgaudit.log_level = 'notice';
-+
-+-- Force parallel execution for testing
-+SET max_parallel_workers_per_gather = 2;
-+SET parallel_tuple_cost = 0;
-+SET parallel_setup_cost = 0;
-+SET min_parallel_table_scan_size = 0;
-+SET min_parallel_index_scan_size = 0;
-+
-+-- Create table with enough data to trigger parallel execution
-+CREATE TABLE parallel_test (id int, data text);
-+INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
-+
-+SELECT count(*) FROM parallel_test;
-+
-+-- Cleanup parallel test
-+DROP TABLE parallel_test;
-+RESET max_parallel_workers_per_gather;
-+RESET parallel_tuple_cost;
-+RESET parallel_setup_cost;
-+RESET min_parallel_table_scan_size;
-+RESET min_parallel_index_scan_size;
-+RESET pgaudit.log;
-+RESET pgaudit.log_client;
-+RESET pgaudit.log_level;
-+
- -- Cleanup
- -- Set client_min_messages up to warning to avoid noise
- SET client_min_messages = 'warning';

compute/patches/pgaudit-parallel_workers-v17.patch

@@ -1,143 +0,0 @@
-commit 8d02e4c6c5e1e8676251b0717a46054267091cb4
-Author: Tristan Partin <tristan.partin@databricks.com>
-Date:   2025-06-23 02:09:31 +0000
-
-    Disable logging in parallel workers
-    
-    When a query uses parallel workers, pgaudit will log the same query for
-    every parallel worker. This is undesireable since it can result in log
-    amplification for queries that use parallel workers.
-    
-    Signed-off-by: Tristan Partin <tristan.partin@databricks.com>
-
-diff --git a/expected/pgaudit.out b/expected/pgaudit.out
-index d696287..4b1059a 100644
---- a/expected/pgaudit.out
-+++ b/expected/pgaudit.out
-@@ -2568,6 +2568,37 @@ DROP SERVER fdw_server;
- NOTICE:  AUDIT: SESSION,11,1,DDL,DROP SERVER,,,DROP SERVER fdw_server,<not logged>
- DROP EXTENSION postgres_fdw;
- NOTICE:  AUDIT: SESSION,12,1,DDL,DROP EXTENSION,,,DROP EXTENSION postgres_fdw,<not logged>
-+--
-+-- Test logging in parallel workers
-+SET pgaudit.log = 'read';
-+SET pgaudit.log_client = on;
-+SET pgaudit.log_level = 'notice';
-+-- Force parallel execution for testing
-+SET max_parallel_workers_per_gather = 2;
-+SET parallel_tuple_cost = 0;
-+SET parallel_setup_cost = 0;
-+SET min_parallel_table_scan_size = 0;
-+SET min_parallel_index_scan_size = 0;
-+-- Create table with enough data to trigger parallel execution
-+CREATE TABLE parallel_test (id int, data text);
-+INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
-+SELECT count(*) FROM parallel_test;
-+NOTICE:  AUDIT: SESSION,13,1,READ,SELECT,,,SELECT count(*) FROM parallel_test,<not logged>
-+ count 
-+-------
-+  1000
-+(1 row)
-+
-+-- Cleanup parallel test
-+DROP TABLE parallel_test;
-+RESET max_parallel_workers_per_gather;
-+RESET parallel_tuple_cost;
-+RESET parallel_setup_cost;
-+RESET min_parallel_table_scan_size;
-+RESET min_parallel_index_scan_size;
-+RESET pgaudit.log;
-+RESET pgaudit.log_client;
-+RESET pgaudit.log_level;
- -- Cleanup
- -- Set client_min_messages up to warning to avoid noise
- SET client_min_messages = 'warning';
-diff --git a/pgaudit.c b/pgaudit.c
-index 1764af1..0e48875 100644
---- a/pgaudit.c
-+++ b/pgaudit.c
-@@ -11,6 +11,7 @@
- #include "postgres.h"
- 
- #include "access/htup_details.h"
-+#include "access/parallel.h"
- #include "access/sysattr.h"
- #include "access/xact.h"
- #include "access/relation.h"
-@@ -1406,7 +1407,7 @@ pgaudit_ExecutorStart_hook(QueryDesc *queryDesc, int eflags)
- {
-     AuditEventStackItem *stackItem = NULL;
- 
--    if (!internalStatement)
-+    if (!internalStatement && !IsParallelWorker())
-     {
-         /* Push the audit event onto the stack */
-         stackItem = stack_push();
-@@ -1489,7 +1490,7 @@ pgaudit_ExecutorCheckPerms_hook(List *rangeTabls, List *permInfos, bool abort)
- 
-     /* Log DML if the audit role is valid or session logging is enabled */
-     if ((auditOid != InvalidOid || auditLogBitmap != 0) &&
--        !IsAbortedTransactionBlockState())
-+        !IsAbortedTransactionBlockState() && !IsParallelWorker())
-     {
-         /* If auditLogRows is on, wait for rows processed to be set */
-         if (auditLogRows && auditEventStack != NULL)
-@@ -1544,7 +1545,7 @@ pgaudit_ExecutorRun_hook(QueryDesc *queryDesc, ScanDirection direction, uint64 c
-     else
-         standard_ExecutorRun(queryDesc, direction, count, execute_once);
- 
--    if (auditLogRows && !internalStatement)
-+    if (auditLogRows && !internalStatement && !IsParallelWorker())
-     {
-         /* Find an item from the stack by the query memory context */
-         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
-@@ -1564,7 +1565,7 @@ pgaudit_ExecutorEnd_hook(QueryDesc *queryDesc)
-     AuditEventStackItem *stackItem = NULL;
-     AuditEventStackItem *auditEventStackFull = NULL;
- 
--    if (auditLogRows && !internalStatement)
-+    if (auditLogRows && !internalStatement && !IsParallelWorker())
-     {
-         /* Find an item from the stack by the query memory context */
-         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
-diff --git a/sql/pgaudit.sql b/sql/pgaudit.sql
-index e161f01..c873098 100644
---- a/sql/pgaudit.sql
-+++ b/sql/pgaudit.sql
-@@ -1637,6 +1637,36 @@ DROP USER MAPPING FOR regress_user1 SERVER fdw_server;
- DROP SERVER fdw_server;
- DROP EXTENSION postgres_fdw;
- 
-+--
-+-- Test logging in parallel workers
-+SET pgaudit.log = 'read';
-+SET pgaudit.log_client = on;
-+SET pgaudit.log_level = 'notice';
-+
-+-- Force parallel execution for testing
-+SET max_parallel_workers_per_gather = 2;
-+SET parallel_tuple_cost = 0;
-+SET parallel_setup_cost = 0;
-+SET min_parallel_table_scan_size = 0;
-+SET min_parallel_index_scan_size = 0;
-+
-+-- Create table with enough data to trigger parallel execution
-+CREATE TABLE parallel_test (id int, data text);
-+INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
-+
-+SELECT count(*) FROM parallel_test;
-+
-+-- Cleanup parallel test
-+DROP TABLE parallel_test;
-+RESET max_parallel_workers_per_gather;
-+RESET parallel_tuple_cost;
-+RESET parallel_setup_cost;
-+RESET min_parallel_table_scan_size;
-+RESET min_parallel_index_scan_size;
-+RESET pgaudit.log;
-+RESET pgaudit.log_client;
-+RESET pgaudit.log_level;
-+
- -- Cleanup
- -- Set client_min_messages up to warning to avoid noise
- SET client_min_messages = 'warning';

compute_tools/Cargo.toml

@@ -27,7 +27,6 @@ fail.workspace = true
 flate2.workspace = true
 futures.workspace = true
 http.workspace = true
-hostname-validator = "1.1"
 indexmap.workspace = true
 itertools.workspace = true
 jsonwebtoken.workspace = true

compute_tools/src/compute.rs

@@ -759,15 +759,10 @@ impl ComputeNode {
         // Configure and start rsyslog for compliance audit logging
         match pspec.spec.audit_log_level {
             ComputeAudit::Hipaa | ComputeAudit::Extended | ComputeAudit::Full => {
-                let remote_tls_endpoint =
-                    std::env::var("AUDIT_LOGGING_TLS_ENDPOINT").unwrap_or("".to_string());
-                let remote_plain_endpoint =
+                let remote_endpoint =
                     std::env::var("AUDIT_LOGGING_ENDPOINT").unwrap_or("".to_string());
-
-                if remote_plain_endpoint.is_empty() && remote_tls_endpoint.is_empty() {
-                    anyhow::bail!(
-                        "AUDIT_LOGGING_ENDPOINT and AUDIT_LOGGING_TLS_ENDPOINT are both empty"
-                    );
+                if remote_endpoint.is_empty() {
+                    anyhow::bail!("AUDIT_LOGGING_ENDPOINT is empty");
                 }
 
                 let log_directory_path = Path::new(&self.params.pgdata).join("log");
@@ -783,8 +778,7 @@ impl ComputeNode {
                     log_directory_path.clone(),
                     endpoint_id,
                     project_id,
-                    &remote_plain_endpoint,
-                    &remote_tls_endpoint,
+                    &remote_endpoint,
                 )?;
 
                 // Launch a background task to clean up the audit logs

compute_tools/src/config_template/compute_audit_rsyslog_template.conf

@@ -10,13 +10,7 @@ input(type="imfile" File="{log_directory}/*.log"
   startmsg.regex="^[[:digit:]]{{4}}-[[:digit:]]{{2}}-[[:digit:]]{{2}} [[:digit:]]{{2}}:[[:digit:]]{{2}}:[[:digit:]]{{2}}.[[:digit:]]{{3}} GMT,")
 
 # the directory to store rsyslog state files
-global(
-  workDirectory="/var/log/rsyslog"
-  DefaultNetstreamDriverCAFile="/etc/ssl/certs/ca-certificates.crt"
-)
-
-# Whether the remote syslog receiver uses tls
-set $.remote_syslog_tls = "{remote_syslog_tls}";
+global(workDirectory="/var/log/rsyslog")
 
 # Construct json, endpoint_id and project_id as additional metadata
 set $.json_log!endpoint_id = "{endpoint_id}";
@@ -27,29 +21,5 @@ set $.json_log!msg = $msg;
 template(name="PgAuditLog" type="string"
     string="<%PRI%>1 %TIMESTAMP:::date-rfc3339% %HOSTNAME% - - - - %$.json_log%")
 
-# Forward to remote syslog receiver (over TLS)
-if ( $syslogtag == 'pgaudit_log' ) then {{
-  if ( $.remote_syslog_tls == 'true' ) then {{
-    action(type="omfwd" target="{remote_syslog_host}" port="{remote_syslog_port}" protocol="tcp"
-      template="PgAuditLog"
-      queue.type="linkedList"
-      queue.size="1000"
-      action.ResumeRetryCount="10"
-      StreamDriver="gtls"
-      StreamDriverMode="1"
-      StreamDriverAuthMode="x509/name"
-      StreamDriverPermittedPeers="{remote_syslog_host}"
-      StreamDriver.CheckExtendedKeyPurpose="on"
-      StreamDriver.PermitExpiredCerts="off"
-    )
-    stop
-  }} else {{
-    action(type="omfwd" target="{remote_syslog_host}" port="{remote_syslog_port}" protocol="tcp"
-      template="PgAuditLog"
-      queue.type="linkedList"
-      queue.size="1000"
-      action.ResumeRetryCount="10"
-    )
-    stop
-  }}
-}}
+# Forward to remote syslog receiver (@@<hostname>:<port>;format
+local5.info @@{remote_endpoint};PgAuditLog

compute_tools/src/rsyslog.rs

@@ -4,10 +4,8 @@ use std::path::Path;
 use std::process::Command;
 use std::time::Duration;
 use std::{fs::OpenOptions, io::Write};
-use url::{Host, Url};
 
 use anyhow::{Context, Result, anyhow};
-use hostname_validator;
 use tracing::{error, info, instrument, warn};
 
 const POSTGRES_LOGS_CONF_PATH: &str = "/etc/rsyslog.d/postgres_logs.conf";
@@ -84,84 +82,18 @@ fn restart_rsyslog() -> Result<()> {
     Ok(())
 }
 
-fn parse_audit_syslog_address(
-    remote_plain_endpoint: &str,
-    remote_tls_endpoint: &str,
-) -> Result<(String, u16, String)> {
-    let tls;
-    let remote_endpoint = if !remote_tls_endpoint.is_empty() {
-        tls = "true".to_string();
-        remote_tls_endpoint
-    } else {
-        tls = "false".to_string();
-        remote_plain_endpoint
-    };
-    // Urlify the remote_endpoint, so parsing can be done with url::Url.
-    let url_str = format!("http://{remote_endpoint}");
-    let url = Url::parse(&url_str).map_err(|err| {
-        anyhow!("Error parsing {remote_endpoint}, expected host:port, got {err:?}")
-    })?;
-
-    let is_valid = url.scheme() == "http"
-        && url.path() == "/"
-        && url.query().is_none()
-        && url.fragment().is_none()
-        && url.username() == ""
-        && url.password().is_none();
-
-    if !is_valid {
-        return Err(anyhow!(
-            "Invalid address format {remote_endpoint}, expected host:port"
-        ));
-    }
-    let host = match url.host() {
-        Some(Host::Domain(h)) if hostname_validator::is_valid(h) => h.to_string(),
-        Some(Host::Ipv4(ip4)) => ip4.to_string(),
-        Some(Host::Ipv6(ip6)) => ip6.to_string(),
-        _ => return Err(anyhow!("Invalid host")),
-    };
-    let port = url
-        .port()
-        .ok_or_else(|| anyhow!("Invalid port in {remote_endpoint}"))?;
-
-    Ok((host, port, tls))
-}
-
-fn generate_audit_rsyslog_config(
-    log_directory: String,
-    endpoint_id: &str,
-    project_id: &str,
-    remote_syslog_host: &str,
-    remote_syslog_port: u16,
-    remote_syslog_tls: &str,
-) -> String {
-    format!(
-        include_str!("config_template/compute_audit_rsyslog_template.conf"),
-        log_directory = log_directory,
-        endpoint_id = endpoint_id,
-        project_id = project_id,
-        remote_syslog_host = remote_syslog_host,
-        remote_syslog_port = remote_syslog_port,
-        remote_syslog_tls = remote_syslog_tls
-    )
-}
-
 pub fn configure_audit_rsyslog(
     log_directory: String,
     endpoint_id: &str,
     project_id: &str,
     remote_endpoint: &str,
-    remote_tls_endpoint: &str,
 ) -> Result<()> {
-    let (remote_syslog_host, remote_syslog_port, remote_syslog_tls) =
-        parse_audit_syslog_address(remote_endpoint, remote_tls_endpoint).unwrap();
-    let config_content = generate_audit_rsyslog_config(
-        log_directory,
-        endpoint_id,
-        project_id,
-        &remote_syslog_host,
-        remote_syslog_port,
-        &remote_syslog_tls,
+    let config_content: String = format!(
+        include_str!("config_template/compute_audit_rsyslog_template.conf"),
+        log_directory = log_directory,
+        endpoint_id = endpoint_id,
+        project_id = project_id,
+        remote_endpoint = remote_endpoint
     );
 
     info!("rsyslog config_content: {}", config_content);
@@ -326,8 +258,6 @@ pub fn launch_pgaudit_gc(log_directory: String) {
 mod tests {
     use crate::rsyslog::PostgresLogsRsyslogConfig;
 
-    use super::{generate_audit_rsyslog_config, parse_audit_syslog_address};
-
     #[test]
     fn test_postgres_logs_config() {
         {
@@ -357,146 +287,4 @@ mod tests {
             assert!(res.is_err());
         }
     }
-
-    #[test]
-    fn test_parse_audit_syslog_address() {
-        {
-            // host:port format (plaintext)
-            let parsed = parse_audit_syslog_address("collector.host.tld:5555", "");
-            assert!(parsed.is_ok());
-            assert_eq!(
-                parsed.unwrap(),
-                (
-                    String::from("collector.host.tld"),
-                    5555,
-                    String::from("false")
-                )
-            );
-        }
-
-        {
-            // host:port format with ipv4 ip address (plaintext)
-            let parsed = parse_audit_syslog_address("10.0.0.1:5555", "");
-            assert!(parsed.is_ok());
-            assert_eq!(
-                parsed.unwrap(),
-                (String::from("10.0.0.1"), 5555, String::from("false"))
-            );
-        }
-
-        {
-            // host:port format with ipv6 ip address (plaintext)
-            let parsed =
-                parse_audit_syslog_address("[7e60:82ed:cb2e:d617:f904:f395:aaca:e252]:5555", "");
-            assert_eq!(
-                parsed.unwrap(),
-                (
-                    String::from("7e60:82ed:cb2e:d617:f904:f395:aaca:e252"),
-                    5555,
-                    String::from("false")
-                )
-            );
-        }
-
-        {
-            // Only TLS host:port defined
-            let parsed = parse_audit_syslog_address("", "tls.host.tld:5556");
-            assert_eq!(
-                parsed.unwrap(),
-                (String::from("tls.host.tld"), 5556, String::from("true"))
-            );
-        }
-
-        {
-            // tls host should take precedence, when both defined
-            let parsed = parse_audit_syslog_address("plaintext.host.tld:5555", "tls.host.tld:5556");
-            assert_eq!(
-                parsed.unwrap(),
-                (String::from("tls.host.tld"), 5556, String::from("true"))
-            );
-        }
-
-        {
-            // host without port (plaintext)
-            let parsed = parse_audit_syslog_address("collector.host.tld", "");
-            assert!(parsed.is_err());
-        }
-
-        {
-            // port without host
-            let parsed = parse_audit_syslog_address(":5555", "");
-            assert!(parsed.is_err());
-        }
-
-        {
-            // valid host with invalid port
-            let parsed = parse_audit_syslog_address("collector.host.tld:90001", "");
-            assert!(parsed.is_err());
-        }
-
-        {
-            // invalid hostname with valid port
-            let parsed = parse_audit_syslog_address("-collector.host.tld:5555", "");
-            assert!(parsed.is_err());
-        }
-
-        {
-            // parse error
-            let parsed = parse_audit_syslog_address("collector.host.tld:::5555", "");
-            assert!(parsed.is_err());
-        }
-    }
-
-    #[test]
-    fn test_generate_audit_rsyslog_config() {
-        {
-            // plaintext version
-            let log_directory = "/tmp/log".to_string();
-            let endpoint_id = "ep-test-endpoint-id";
-            let project_id = "test-project-id";
-            let remote_syslog_host = "collector.host.tld";
-            let remote_syslog_port = 5555;
-            let remote_syslog_tls = "false";
-
-            let conf_str = generate_audit_rsyslog_config(
-                log_directory,
-                endpoint_id,
-                project_id,
-                remote_syslog_host,
-                remote_syslog_port,
-                remote_syslog_tls,
-            );
-
-            assert!(conf_str.contains(r#"set $.remote_syslog_tls = "false";"#));
-            assert!(conf_str.contains(r#"type="omfwd""#));
-            assert!(conf_str.contains(r#"target="collector.host.tld""#));
-            assert!(conf_str.contains(r#"port="5555""#));
-            assert!(conf_str.contains(r#"StreamDriverPermittedPeers="collector.host.tld""#));
-        }
-
-        {
-            // TLS version
-            let log_directory = "/tmp/log".to_string();
-            let endpoint_id = "ep-test-endpoint-id";
-            let project_id = "test-project-id";
-            let remote_syslog_host = "collector.host.tld";
-            let remote_syslog_port = 5556;
-            let remote_syslog_tls = "true";
-
-            let conf_str = generate_audit_rsyslog_config(
-                log_directory,
-                endpoint_id,
-                project_id,
-                remote_syslog_host,
-                remote_syslog_port,
-                remote_syslog_tls,
-            );
-
-            assert!(conf_str.contains(r#"set $.remote_syslog_tls = "true";"#));
-            assert!(conf_str.contains(r#"type="omfwd""#));
-            assert!(conf_str.contains(r#"target="collector.host.tld""#));
-            assert!(conf_str.contains(r#"port="5556""#));
-            assert!(conf_str.contains(r#"StreamDriverPermittedPeers="collector.host.tld""#));
-        }
-    }
 }

docs/rfcs/044-feature-flag.md

@@ -1,179 +0,0 @@
-# Storage Feature Flags
-
-In this RFC, we will describe how we will implement per-tenant feature flags.
-
-## PostHog as Feature Flag Service
-
-Before we start, let's talk about how current feature flag services work. PostHog is the feature flag service we are currently using across multiple user-facing components in the company. PostHog has two modes of operation: HTTP evaluation and server-side local evaluation.
-
-Let's assume we have a storage feature flag called gc-compaction and we want to roll it out to scale-tier users with resident size >= 10GB and <= 100GB.
-
-### Define User Profiles
-
-The first step is to synchronize our user profiles to the PostHog service. We can simply assume that each tenant is a user in PostHog. Each user profile has some properties associated with it. In our case, it will be: plan type (free, scale, enterprise, etc); resident size (in bytes); primary pageserver (string); region (string).
-
-### Define Feature Flags
-
-We would create a feature flag called gc-compaction in PostHog with 4 variants: disabled, stage-1, stage-2, fully-enabled. We will flip the feature flags from disabled to fully-enabled stage by stage for some percentage of our users.
-
-### Option 1: HTTP Evaluation Mode
-
-When using PostHog's HTTP evaluation mode, the client will make request to the PostHog service, asking for the value of a feature flag for a specific user.
-
-* Control plane will report the plan type to PostHog each time it attaches a tenant to the storcon or when the user upgrades/downgrades. It calls the PostHog profile API to associate tenant ID with the plan type. Assume we have X active tenants and such attach or plan change event happens each week, that would be 4X profile update requests per month.
-* Pageservers will report the resident size and the primary pageserver to the PostHog service. Assume we report resident size every 24 hours, that would be 30X requests per month.
-* Each tenant will request the state of the feature flag every 1 hour, that's 720X requests per month.
-* The Rust client would be easy to implement as we only need to call the `/decide` API on PostHog.
-
-Using the HTTP evaluation mode we will issue 754X requests a month.
-
-### Option 2: Local Evaluation Mode
-
-When using PostHog's HTTP evaluation mode, the client (usually the server in a browser/server architecture) will poll the feature flag configuration every 30s (default in the Python client) from PostHog. Such configuration contains data like:
-
-<details>
-
-<summary>Example JSON response from the PostHog local evaluation API</summary>
-
-```
-[
-    {
-        "id": 1,
-        "name": "Beta Feature",
-        "key": "person-flag",
-        "is_simple_flag": True,
-        "active": True,
-        "filters": {
-            "groups": [
-                {
-                    "properties": [
-                        {
-                            "key": "location",
-                            "operator": "exact",
-                            "value": ["Straße"],
-                            "type": "person",
-                        }
-                    ],
-                    "rollout_percentage": 100,
-                },
-                {
-                    "properties": [
-                        {
-                            "key": "star",
-                            "operator": "exact",
-                            "value": ["ſun"],
-                            "type": "person",
-                        }
-                    ],
-                    "rollout_percentage": 100,
-                },
-            ],
-        },
-    }
-]
-```
-
-</details>
-
-Note that the API only contains information like "under what condition => rollout percentage". The user is responsible to provide the properties required to the client for local evaluation, and the PostHog service (web UI) cannot know if a feature is enabled for the tenant or not until the client uses the `capture` API to report the result back. To control the rollout percentage, the user ID gets mapped to a float number in `[0, 1)` on a consistent hash ring. All values <= the percentage will get the feature enabled or set to the desired value.
-
-To use the local evaluation mode, the system needs:
-
-* Assume each pageserver will poll PostHog for the local evaluation JSON every 5 minutes (instead of the 30s default as it's too frequent). That's 8640Y per month, Y is the number of pageservers. Local evaluation requests cost 10x more than the normal decide request, so that's 86400Y request units to bill.
-* Storcon needs to store the plan type in the database and pass that information to the pageserver when attaching the tenant.
-* Storcon also needs to update PostHog with the active tenants, for example, when the tenant gets detached/attached. Assume each active tenant gets detached/attached every week, that would be 4X requests per month.
-* We do not need to update bill type or resident size to PostHog as all these are evaluated locally.
-* After each local evaluation of the feature flag, we need to call PostHog's capture event API to update the result of the evaluation that the feature is enabled. We can do this when the flag gets changed compared with the last cached state in memory. That would be at least 4X (assume we do deployment every week so the cache gets cleared) and maybe an additional multiplifier of 10 assume we have 10 active features.
-
-In this case, we will issue 86400Y + 40X requests per month.
-
-Assume X = 1,000,000 and Y = 100,
-
-|   | HTTP Evaluation  | Local Evaluation  |
-|---|---|---|
-| Latency of propagating the conditions/properties for feature flag  | 24 hours  | available locally  |
-| Latency of applying the feature flag  | 1 hour  | 5 minutes  |
-| Can properties be reported from different services |  Yes |  No  |
-| Do we need to sync billing info etc to pageserver |  No |  Yes  |
-| Cost | 75400$ / month | 4864$ / month |
-
-# Our Solution
-
-We will use PostHog _only_ as an UI to configure the feature flags. Whether a feature is enabled or not can only be queried through storcon/pageserver instead of using the PostHog UI. (We could report it back to PostHog via `capture_event` but it costs $$$.) This allows us to ramp up the feature flag functionality fast at first. At the same time, it would also give us the option to migrate to our own solution once we want to have more properties and more complex evaluation rules in our system.
-
-* We will create several fake users (tenants) in PostHog that contains all the properties we will use for evaluating a feature flag (i.e., resident size, billing type, pageserver id, etc.)
-* We will use PostHog's local evaluation API to poll the configuration of the feature flags and evaluate them locally on each of the pageserver.
-* The evaluation result will not be reported back to PostHog.
-* Storcon needs to pull some information from cplane database.
-* To know if a feature is currently enabled or not, we need to call the storcon/pageserver API; and we won't be able to know if a feature has been enabled on a tenant before easily: we need to look at the Grafana logs.
-
-We only need to pay for the 86400Y local evaluation requests (that would be setting Y=0 in solution 2 => $864/month, and even less if we proxy it through storcon).
-
-## Implementation
-
-* Pageserver: implement a PostHog local evaluation client. The client will be shared across all tenants on the pageserver with a single API: `evaluate(tenant_id, feature_flag, properties) -> json`.
-* Storcon: if we need plan type as the evaluation condition, pull it from cplane database.
-* Storcon/Pageserver: implement an HTTP API `:tenant_id/feature/:feature` to retrieve the current feature flag status.
-* Storcon/Pageserver: a loop to update the feature flag spec on both storcon and pageserver. Pageserver loop will only be activated if storcon does not push the specs to the pageserver.
-
-## Difference from Tenant Config
-
-* Feature flags can be modified by percentage, and the default config for each feature flag can be modified in UI without going through the release process.
-* Feature flags are more flexible and won't be persisted anywhere and will be passed as plain JSON over the wire so that do not need to handle backward/forward compatibility as in tenant config.
-* The expectation of tenant config is that once we add a flag we cannot remove it (or it will be hard to remove), but feature flags are more flexible.
-
-# Final Implementation
-
-* We added a new crate `posthog_lite_client` that supports local feature evaluations.
-* We set up two projects "Storage (staging)" and "Storage (production)" in the PostHog console.
-* Each pageserver reports 10 fake tenants to PostHog so that we can get all combinations of regions (and other properties) in the PostHog UI.
-* Supported properties: AZ, neon_region, pageserver, tenant_id.
-* You may use "Pageserver Feature Flags" dashboard to see the evaluation status.
-* The feature flag spec is polled on storcon every 30s (in each of the region) and storcon will propagate the spec to the pageservers.
-* The pageserver housekeeping loop updates the tenant-specific properties (e.g., remote size) for evaluation.
-
-Each tenant has a `feature_resolver` object. After you add a feature flag in the PostHog console, you can retrieve it with:
-
-```rust
-// Boolean flag
-self
-    .feature_resolver
-    .evaluate_boolean("flag")
-    .is_ok()
-// Multivariate flag
-self
-    .feature_resolver
-    .evaluate_multivariate("gc-comapction-strategy")
-    .ok();
-```
-
-The user needs to handle the case where the evaluation result is an error. This can occur in a variety of cases:
-
-* During the pageserver start, the feature flag spec has not been retrieved.
-* No condition group is matched.
-* The feature flag spec contains an operand/operation not supported by the lite PostHog library.
-
-For boolean flags, the return value is `Result<(), Error>`. `Ok(())` means the flag is evaluated to true. Otherwise,
-there is either an error in evaluation or it does not match any groups.
-
-For multivariate flags, the return value is `Result<String, Error>`. `Ok(variant)` indicates the flag is evaluated
-to a variant. Otherwise, there is either an error in evaluation or it does not match any groups.
-
-The evaluation logic is documented in the PostHog lite library. It compares the consistent hash of a flag key + tenant_id
-with the rollout percentage and determines which tenant to roll out a specific feature.
-
-Users can use the feature flag evaluation API to get the flag evaluation result of a specific tenant for debugging purposes.
-
-```
-curl http://localhost:9898/v1/tenant/:tenant_id/feature_flag?flag=:key&as=multivariate/boolean"
-```
-
-By default, the storcon pushes the feature flag specs to the pageservers every 30 seconds, which means that a change in feature flag in the
-PostHog UI will propagate to the pageservers within 30 seconds.
-
-# Future Works
-
-* Support dynamic tenant properties like logical size as the evaluation condition.
-* Support properties like `plan_type` (needs cplane to pass it down).
-* Report feature flag evaluation result back to PostHog (if the cost is okay).
-* Fast feature flag evaluation cache on critical paths (e.g., cache a feature flag result in `AtomicBool` and use it on the read path).

docs/rfcs/2025-03-17-compute-prewarm.md

@@ -1,399 +0,0 @@
-# Compute rolling restart with prewarm
-
-Created on 2025-03-17
-Implemented on _TBD_
-Author: Alexey Kondratov (@ololobus)
-
-## Summary
-
-This RFC describes an approach to reduce performance degradation due to missing caches after compute node restart, i.e.:
-
-1. Rolling restart of the running instance via 'warm' replica.
-2. Auto-prewarm compute caches after unplanned restart or scale-to-zero.
-
-## Motivation
-
-Neon currently implements several features that guarantee high uptime of compute nodes:
-
-1. Storage high-availability (HA), i.e. each tenant shard has a secondary pageserver location, so we can quickly switch over compute to it in case of primary pageserver failure.
-2. Fast compute provisioning, i.e. we have a fleet of pre-created empty computes, that are ready to serve workload, so restarting unresponsive compute is very fast.
-3. Preemptive NeonVM compute provisioning in case of k8s node unavailability.
-
-This helps us to be well-within the uptime SLO of 99.95% most of the time. Problems begin when we go up to multi-TB workloads and 32-64 CU computes.
-During restart, compute loses all caches: LFC, shared buffers, file system cache. Depending on the workload, it can take a lot of time to warm up the caches,
-so that performance could be degraded and might be even unacceptable for certain workloads. The latter means that although current approach works well for small to
-medium workloads, we still have to do some additional work to avoid performance degradation after restart of large instances.
-
-## Non Goals
-
-- Details of the persistence storage for prewarm data are out of scope, there is a separate RFC for that: <https://github.com/neondatabase/neon/pull/9661>.
-- Complete compute/Postgres HA setup and flow. Although it was originally in scope of this RFC, during preliminary research it appeared to be a rabbit hole, so it's worth of a separate RFC.
-- Low-level implementation details for Postgres replica-to-primary promotion. There are a lot of things to think and care about: how to start walproposer, [logical replication failover](https://www.postgresql.org/docs/current/logical-replication-failover.html), and so on, but it's worth of at least a separate one-pager design document if not RFC.
-
-## Impacted components
-
-Postgres, compute_ctl, Control plane, Endpoint storage for unlogged storage of compute files.
-For the latter, we will need to implement a uniform abstraction layer on top of S3, ABS, etc., but
-S3 is used in text interchangeably with 'endpoint storage' for simplicity.
-
-## Proposed implementation
-
-### compute_ctl spec changes and auto-prewarm
-
-We are going to extend the current compute spec with the following attributes
-
-```rust
-struct ComputeSpec {
-    /// [All existing attributes]
-    ...
-    /// Whether to do auto-prewarm at start or not.
-    /// Default to `false`.
-    pub lfc_auto_prewarm: bool
-    /// Interval in seconds between automatic dumps of
-    /// LFC state into S3. Default `None`, which means 'off'.
-    pub lfc_dump_interval_sec: Option<i32>
-}
-```
-
-When `lfc_dump_interval_sec` is set to `N`, `compute_ctl` will periodically dump the LFC state
-and store it in S3, so that it could be used either for auto-prewarm after restart or by replica
-during the rolling restart. For enabling periodic dumping, we should consider the following value
-`lfc_dump_interval_sec=300` (5 minutes), same as in the upstream's `pg_prewarm.autoprewarm_interval`.
-
-When `lfc_auto_prewarm` is set to `true`, `compute_ctl` will start prewarming the LFC upon restart
-iif some of the previous states is present in S3.
-
-### compute_ctl API
-
-1. `POST /store_lfc_state` -- dump LFC state using Postgres SQL interface and store result in S3.
-    This has to be a blocking call, i.e. it will return only after the state is stored in S3.
-    If there is any concurrent request in progress, we should return `429 Too Many Requests`,
-    and let the caller to retry.
-
-2. `GET /dump_lfc_state` -- dump LFC state using Postgres SQL interface and return it as is
-    in text format suitable for the future restore/prewarm. This API is not strictly needed at
-    the end state, but could be useful for a faster prototyping of a complete rolling restart flow
-    with prewarm, as it doesn't require persistent for LFC state storage.
-
-3. `POST /restore_lfc_state` -- restore/prewarm LFC state with request
-
-    ```yaml
-    RestoreLFCStateRequest:
-      oneOf:
-        - type: object
-          required:
-            - lfc_state
-          properties:
-            lfc_state:
-              type: string
-              description: Raw LFC content dumped with GET `/dump_lfc_state`
-        - type: object
-          required:
-            - lfc_cache_key
-          properties:
-            lfc_cache_key:
-              type: string
-              description: |
-                endpoint_id of the source endpoint on the same branch
-                to use as a 'donor' for LFC content. Compute will look up
-                LFC content dump in S3 using this key and do prewarm.
-    ```
-
-    where `lfc_state` and `lfc_cache_key` are mutually exclusive.
-
-    The actual prewarming will happen asynchronously, so the caller need to check the
-    prewarm status using the compute's standard `GET /status` API.
-
-4. `GET /status` -- extend existing API with following attributes
-
-    ```rust
-    struct ComputeStatusResponse {
-        // [All existing attributes]
-        ...
-        pub prewarm_state: PrewarmState
-    }
-
-    /// Compute prewarm state. Will be stored in the shared Compute state
-    /// in compute_ctl
-    struct PrewarmState {
-        pub status: PrewarmStatus
-        /// Total number of pages to prewarm
-        pub pages_total: i64
-        /// Number of pages prewarmed so far
-        pub pages_processed: i64
-        /// Optional prewarm error
-        pub error: Option<String>
-    }
-
-    pub enum PrewarmStatus {
-        /// Prewarming was never requested on this compute
-        Off,
-        /// Prewarming was requested, but not started yet
-        Pending,
-        /// Prewarming is in progress. The caller should follow
-        /// `PrewarmState::progress`.
-        InProgress,
-        /// Prewarming has been successfully completed
-        Completed,
-        /// Prewarming failed. The caller should look at
-        /// `PrewarmState::error` for the reason.
-        Failed,
-        /// It is intended to be used by auto-prewarm if none of
-        /// the previous LFC states is available in S3.
-        /// This is a distinct state from the `Failed` because
-        /// technically it's not a failure and could happen if
-        /// compute was restart before it dumped anything into S3,
-        /// or just after the initial rollout of the feature.
-        Skipped,
-    }
-    ```
-
-5. `POST /promote` -- this is a **blocking** API call to promote compute replica into primary.
-    This API should be very similar to the existing `POST /configure` API, i.e. accept the
-    spec (primary spec, because originally compute was started as replica). It's a distinct
-    API method because semantics and response codes are different:
-
-    - If promotion is done successfully, it will return `200 OK`.
-    - If compute is already primary, the call will be no-op and `compute_ctl`
-      will return `412 Precondition Failed`.
-    - If, for some reason, second request reaches compute that is in progress of promotion,
-      it will respond with `429 Too Many Requests`.
-    - If compute hit any permanent failure during promotion `500 Internal Server Error`
-      will be returned.
-
-### Control plane operations
-
-The complete flow will be present as a sequence diagram in the next section, but here
-we just want to list some important steps that have to be done by control plane during
-the rolling restart via warm replica, but without much of low-level implementation details.
-
-1. Register the 'intent' of the instance restart, but not yet interrupt any workload at
-    primary and also accept new connections. This may require some endpoint state machine
-    changes, e.g. introduction of the `pending_restart` state. Being in this state also
-    **mustn't prevent any other operations except restart**: suspend, live-reconfiguration
-    (e.g. due to notify-attach call from the storage controller), deletion.
-
-2. Start new replica compute on the same timeline and start prewarming it. This process
-    may take quite a while, so the same concurrency considerations as in 1. should be applied
-    here as well.
-
-3. When warm replica is ready, control plane should:
-
-    3.1. Terminate the primary compute. Starting from here, **this is a critical section**,
-        if anything goes off, the only option is to start the primary normally and proceed
-        with auto-prewarm.
-
-    3.2. Send cache invalidation message to all proxies, notifying them that all new connections
-        should request and wait for the new connection details. At this stage, proxy has to also
-        drop any existing connections to the old primary, so they didn't do stale reads.
-
-    3.3. Attach warm replica compute to the primary endpoint inside control plane metadata
-        database.
-
-    3.4. Promote replica to primary.
-
-    3.5. When everything is done, finalize the endpoint state to be just `active`.
-
-### Complete rolling restart flow
-
-```mermaid
-  sequenceDiagram
-
-  autonumber
-
-  participant proxy as Neon proxy
-
-  participant cplane as Control plane
-
-  participant primary as Compute (primary)
-  box Compute (replica)
-    participant ctl as compute_ctl
-    participant pg as Postgres
-  end
-
-  box Endpoint unlogged storage
-    participant s3proxy as Endpoint storage service
-    participant s3 as S3/ABS/etc.
-  end
-
-
-  cplane ->> primary: POST /store_lfc_state
-  primary -->> cplane: 200 OK
-
-  cplane ->> ctl: POST /restore_lfc_state
-  activate ctl
-  ctl -->> cplane: 202 Accepted
-
-  activate cplane
-  cplane ->> ctl: GET /status: poll prewarm status
-  ctl ->> s3proxy: GET /read_file
-  s3proxy ->> s3: read file
-  s3 -->> s3proxy: file content
-  s3proxy -->> ctl: 200 OK: file content
-
-  proxy ->> cplane: GET /proxy_wake_compute
-  cplane -->> proxy: 200 OK: old primary conninfo
-
-  ctl ->> pg: prewarm LFC
-  activate pg
-  pg -->> ctl: prewarm is completed
-  deactivate pg
-
-  ctl -->> cplane: 200 OK: prewarm is completed
-  deactivate ctl
-  deactivate cplane
-
-  cplane -->> cplane: reassign replica compute to endpoint,<br>start terminating the old primary compute
-  activate cplane
-  cplane ->> proxy: invalidate caches
-
-  proxy ->> cplane: GET /proxy_wake_compute
-
-  cplane -x primary: POST /terminate
-  primary -->> cplane: 200 OK
-  note over primary: old primary<br>compute terminated
-
-  cplane ->> ctl: POST /promote
-  activate ctl
-  ctl ->> pg: pg_ctl promote
-  activate pg
-  pg -->> ctl: done
-  deactivate pg
-  ctl -->> cplane: 200 OK
-  deactivate ctl
-
-  cplane -->> cplane: finalize operation
-  cplane -->> proxy: 200 OK: new primary conninfo
-  deactivate cplane
-```
-
-### Network bandwidth and prewarm speed
-
-It's currently known that pageserver can sustain about 3000 RPS per shard for a few running computes.
-Large tenants are usually split into 8 shards, so the final formula may look like this:
-
-```text
-8 shards * 3000 RPS * 8 KB =~ 190 MB/s
-```
-
-so depending on the LFC size, prewarming will take at least:
-
-- ~5s for 1 GB
-- ~50s for 10 GB
-- ~5m for 100 GB
-- \>1h for 1 TB
-
-In total, one pageserver is normally capped by 30k RPS, so it obviously can't sustain many computes
-doing prewarm at the same time. Later, we may need an additional mechanism for computes to throttle
-the prewarming requests gracefully.
-
-### Reliability, failure modes and corner cases
-
-We consider following failures while implementing this RFC:
-
-1. Compute got interrupted/crashed/restarted during prewarm. The caller -- control plane -- should
-    detect that and start prewarm from the beginning.
-
-2. Control plane promotion request timed out or hit network issues. If it never reached the
-    compute, control plane should just repeat it. If it did reach the compute, then during
-    retry control plane can hit `409` as previous request triggered the promotion already.
-    In this case, control plane need to retry until either `200` or
-    permanent error `500` is returned.
-
-3. Compute got interrupted/crashed/restarted during promotion. At restart it will ask for
-    a spec from control plane, and its content should signal compute to start as **primary**,
-    so it's expected that control plane will continue polling for certain period of time and
-    will discover that compute is ready to accept connections if restart is fast enough.
-
-4. Any other unexpected failure or timeout during prewarming. This **failure mustn't be fatal**,
-    control plane has to report failure, terminate replica and keep primary running.
-
-5. Any other unexpected failure or timeout during promotion. Unfortunately, at this moment
-    we already have the primary node stopped, so the only option is to start primary again
-    and proceed with auto-prewarm.
-
-6. Any unexpected failure during auto-prewarm. This **failure mustn't be fatal**,
-    `compute_ctl` has to report the failure, but do not crash the compute.
-
-7. Control plane failed to confirm that old primary has terminated. This can happen, especially
-    in the future HA setup. In this case, control plane has to ensure that it sent VM deletion
-    and pod termination requests to k8s, so long-term we do not have two running primaries
-    on the same timeline.
-
-### Security implications
-
-There are two security implications to consider:
-
-1. Access to `compute_ctl` API. It has to be accessible from the outside of compute, so all
-    new API methods have to be exposed on the **external** HTTP port and **must** be authenticated
-    with JWT.
-
-2. Read/write only your own LFC state data in S3. Although it's not really a security concern,
-    since LFC state is just a mapping of blocks present in LFC at certain moment in time;
-    it still has to be highly restricted, so that i) only computes on the same timeline can
-    read S3 state; ii) each compute can only write to the path that contains it's `endpoint_id`.
-    Both of this must be validated by Endpoint storage service using the JWT token provided by `compute_ctl`.
-
-### Unresolved questions
-
-#### Billing, metrics and monitoring
-
-Currently, we only label computes with `endpoint_id` after attaching them to the endpoint.
-In this proposal, this means that temporary replica will remain unlabelled until it's promoted
-to primary. We can also hide it from users in the control plane API, but what to do with
-billing and monitoring is still unclear.
-
-We can probably mark it as 'billable' and tag with `project_id`, so it will be billed, but
-not interfere in any way with the current primary monitoring.
-
-Another thing to consider is how logs and metrics export will switch to the new compute.
-It's expected that OpenTelemetry collector will auto-discover the new compute and start
-scraping metrics from it.
-
-#### Auto-prewarm
-
-It's still an open question whether we need auto-prewarm at all. The author's gut-feeling is
-that yes, we need it, but might be not for all workloads, so it could end up exposed as a
-user-controllable knob on the endpoint. There are two arguments for that:
-
-1. Auto-prewarm existing in upstream's `pg_prewarm`, _probably for a reason_.
-
-2. There are still could be 2 flows when we cannot perform the rolling restart via the warm
-    replica: i) any failure or interruption during promotion; ii) wake up after scale-to-zero.
-    The latter might be challenged as well, i.e. one can argue that auto-prewarm may and will
-    compete with user-workload for storage resources. This is correct, but it might as well
-    reduce the time to get warm LFC and good performance.
-
-#### Low-level details of the replica promotion
-
-There are many things to consider here, but three items just off the top of my head:
-
-1. How to properly start the `walproposer` inside Postgres.
-
-2. What to do with logical replication. Currently, we do not include logical replication slots
-    inside basebackup, because nobody advances them at replica, so they just prevent the WAL
-    deletion. Yet, we do need to have them at primary after promotion. Starting with Postgres 17,
-    there is a new feature called
-    [logical replication failover](https://www.postgresql.org/docs/current/logical-replication-failover.html)
-    and `synchronized_standby_slots` setting, but we need a plan for the older versions. Should we
-    request a new basebackup during promotion?
-
-3. How do we guarantee that replica will receive all the latest WAL from safekeepers? Do some
-    'shallow' version of sync safekeepers without data copying? Or just a standard version of
-    sync safekeepers?
-
-## Alternative implementation
-
-The proposal already assumes one of the alternatives -- do not have any persistent storage for
-LFC state. This is possible to implement faster with the proposed API, but it means that
-we do not implement auto-prewarm yet.
-
-## Definition of Done
-
-At the end of implementing this RFC we should have two high-level settings that enable:
-
-1. Auto-prewarm of user computes upon restart.
-2. Perform primary compute restart via the warm replica promotion.
-
-It also has to be decided what's the criteria for enabling one or both of these flows for
-certain clients.

pageserver/src/consumption_metrics/upload.rs

@@ -99,7 +99,7 @@ pub(super) async fn upload_metrics_bucket(
 
     // Compose object path
     let datetime: DateTime<Utc> = SystemTime::now().into();
-    let ts_prefix = datetime.format("year=%Y/month=%m/day=%d/hour=%H/%H:%M:%SZ");
+    let ts_prefix = datetime.format("year=%Y/month=%m/day=%d/%H:%M:%SZ");
     let path = RemotePath::from_string(&format!("{ts_prefix}_{node_id}.ndjson.gz"))?;
 
     // Set up a gzip writer into a buffer
@@ -109,7 +109,7 @@ pub(super) async fn upload_metrics_bucket(
 
     // Serialize and write into compressed buffer
     let started_at = std::time::Instant::now();
-    for res in serialize_in_chunks_ndjson(CHUNK_SIZE, metrics, idempotency_keys) {
+    for res in serialize_in_chunks(CHUNK_SIZE, metrics, idempotency_keys) {
         let (_chunk, body) = res?;
         gzip_writer.write_all(&body).await?;
     }
@@ -216,86 +216,6 @@ fn serialize_in_chunks<'a>(
     }
 }
 
-/// Serializes the input metrics as NDJSON in chunks of chunk_size. Each event
-/// is serialized as a separate JSON object on its own line. The provided
-/// idempotency keys are injected into the corresponding metric events (reused
-/// across different metrics sinks), and must have the same length as input.
-fn serialize_in_chunks_ndjson<'a>(
-    chunk_size: usize,
-    input: &'a [NewRawMetric],
-    idempotency_keys: &'a [IdempotencyKey<'a>],
-) -> impl ExactSizeIterator<Item = Result<(&'a [NewRawMetric], bytes::Bytes), serde_json::Error>> + 'a
-{
-    use bytes::BufMut;
-
-    assert_eq!(input.len(), idempotency_keys.len());
-
-    struct Iter<'a> {
-        inner: std::slice::Chunks<'a, NewRawMetric>,
-        idempotency_keys: std::slice::Iter<'a, IdempotencyKey<'a>>,
-        chunk_size: usize,
-
-        // write to a BytesMut so that we can cheaply clone the frozen Bytes for retries
-        buffer: bytes::BytesMut,
-        // chunk amount of events are reused to produce the serialized document
-        scratch: Vec<Event<Ids, Name>>,
-    }
-
-    impl<'a> Iterator for Iter<'a> {
-        type Item = Result<(&'a [NewRawMetric], bytes::Bytes), serde_json::Error>;
-
-        fn next(&mut self) -> Option<Self::Item> {
-            let chunk = self.inner.next()?;
-
-            if self.scratch.is_empty() {
-                // first round: create events with N strings
-                self.scratch.extend(
-                    chunk
-                        .iter()
-                        .zip(&mut self.idempotency_keys)
-                        .map(|(raw_metric, key)| raw_metric.as_event(key)),
-                );
-            } else {
-                // next rounds: update_in_place to reuse allocations
-                assert_eq!(self.scratch.len(), self.chunk_size);
-                itertools::izip!(self.scratch.iter_mut(), chunk, &mut self.idempotency_keys)
-                    .for_each(|(slot, raw_metric, key)| raw_metric.update_in_place(slot, key));
-            }
-
-            // Serialize each event as NDJSON (one JSON object per line)
-            for event in self.scratch[..chunk.len()].iter() {
-                let res = serde_json::to_writer((&mut self.buffer).writer(), event);
-                if let Err(e) = res {
-                    return Some(Err(e));
-                }
-                // Add newline after each event to follow NDJSON format
-                self.buffer.put_u8(b'\n');
-            }
-
-            Some(Ok((chunk, self.buffer.split().freeze())))
-        }
-
-        fn size_hint(&self) -> (usize, Option<usize>) {
-            self.inner.size_hint()
-        }
-    }
-
-    impl ExactSizeIterator for Iter<'_> {}
-
-    let buffer = bytes::BytesMut::new();
-    let inner = input.chunks(chunk_size);
-    let idempotency_keys = idempotency_keys.iter();
-    let scratch = Vec::new();
-
-    Iter {
-        inner,
-        idempotency_keys,
-        chunk_size,
-        buffer,
-        scratch,
-    }
-}
-
 trait RawMetricExt {
     fn as_event(&self, key: &IdempotencyKey<'_>) -> Event<Ids, Name>;
     fn update_in_place(&self, event: &mut Event<Ids, Name>, key: &IdempotencyKey<'_>);
@@ -559,43 +479,6 @@ mod tests {
         }
     }
 
-    #[test]
-    fn chunked_serialization_ndjson() {
-        let examples = metric_samples();
-        assert!(examples.len() > 1);
-
-        let now = Utc::now();
-        let idempotency_keys = (0..examples.len())
-            .map(|i| FixedGen::new(now, "1", i as u16).generate())
-            .collect::<Vec<_>>();
-
-        // Parse NDJSON format - each line is a separate JSON object
-        let parse_ndjson = |body: &[u8]| -> Vec<Event<Ids, Name>> {
-            let body_str = std::str::from_utf8(body).unwrap();
-            body_str
-                .trim_end_matches('\n')
-                .lines()
-                .filter(|line| !line.is_empty())
-                .map(|line| serde_json::from_str::<Event<Ids, Name>>(line).unwrap())
-                .collect()
-        };
-
-        let correct = serialize_in_chunks_ndjson(examples.len(), &examples, &idempotency_keys)
-            .map(|res| res.unwrap().1)
-            .flat_map(|body| parse_ndjson(&body))
-            .collect::<Vec<_>>();
-
-        for chunk_size in 1..examples.len() {
-            let actual = serialize_in_chunks_ndjson(chunk_size, &examples, &idempotency_keys)
-                .map(|res| res.unwrap().1)
-                .flat_map(|body| parse_ndjson(&body))
-                .collect::<Vec<_>>();
-
-            // if these are equal, it means that multi-chunking version works as well
-            assert_eq!(correct, actual);
-        }
-    }
-
     #[derive(Clone, Copy)]
     struct FixedGen<'a>(chrono::DateTime<chrono::Utc>, &'a str, u16);
 

pageserver/src/feature_resolver.rs

@@ -6,13 +6,12 @@ use posthog_client_lite::{
     CaptureEvent, FeatureResolverBackgroundLoop, PostHogEvaluationError,
     PostHogFlagFilterPropertyValue,
 };
-use rand::Rng;
 use remote_storage::RemoteStorageKind;
 use serde_json::json;
 use tokio_util::sync::CancellationToken;
 use utils::id::TenantId;
 
-use crate::{config::PageServerConf, metrics::FEATURE_FLAG_EVALUATION, tenant::TenantShard};
+use crate::{config::PageServerConf, metrics::FEATURE_FLAG_EVALUATION};
 
 const DEFAULT_POSTHOG_REFRESH_INTERVAL: Duration = Duration::from_secs(600);
 
@@ -139,7 +138,6 @@ impl FeatureResolver {
                 }
                 Arc::new(properties)
             };
-
             let fake_tenants = {
                 let mut tenants = Vec::new();
                 for i in 0..10 {
@@ -149,16 +147,9 @@ impl FeatureResolver {
                         conf.id,
                         i
                     );
-
-                    let tenant_properties = PerTenantProperties {
-                        remote_size_mb: Some(rand::thread_rng().gen_range(100.0..1000000.00)),
-                    }
-                    .into_posthog_properties();
-
                     let properties = Self::collect_properties_inner(
                         distinct_id.clone(),
                         Some(&internal_properties),
-                        &tenant_properties,
                     );
                     tenants.push(CaptureEvent {
                         event: "initial_tenant_report".to_string(),
@@ -192,7 +183,6 @@ impl FeatureResolver {
     fn collect_properties_inner(
         tenant_id: String,
         internal_properties: Option<&HashMap<String, PostHogFlagFilterPropertyValue>>,
-        tenant_properties: &HashMap<String, PostHogFlagFilterPropertyValue>,
     ) -> HashMap<String, PostHogFlagFilterPropertyValue> {
         let mut properties = HashMap::new();
         if let Some(internal_properties) = internal_properties {
@@ -204,9 +194,6 @@ impl FeatureResolver {
             "tenant_id".to_string(),
             PostHogFlagFilterPropertyValue::String(tenant_id),
         );
-        for (key, value) in tenant_properties.iter() {
-            properties.insert(key.clone(), value.clone());
-        }
         properties
     }
 
@@ -214,13 +201,8 @@ impl FeatureResolver {
     pub(crate) fn collect_properties(
         &self,
         tenant_id: TenantId,
-        tenant_properties: &HashMap<String, PostHogFlagFilterPropertyValue>,
     ) -> HashMap<String, PostHogFlagFilterPropertyValue> {
-        Self::collect_properties_inner(
-            tenant_id.to_string(),
-            self.internal_properties.as_deref(),
-            tenant_properties,
-        )
+        Self::collect_properties_inner(tenant_id.to_string(), self.internal_properties.as_deref())
     }
 
     /// Evaluate a multivariate feature flag. Currently, we do not support any properties.
@@ -232,7 +214,6 @@ impl FeatureResolver {
         &self,
         flag_key: &str,
         tenant_id: TenantId,
-        tenant_properties: &HashMap<String, PostHogFlagFilterPropertyValue>,
     ) -> Result<String, PostHogEvaluationError> {
         let force_overrides = self.force_overrides_for_testing.load();
         if let Some(value) = force_overrides.get(flag_key) {
@@ -243,7 +224,7 @@ impl FeatureResolver {
             let res = inner.feature_store().evaluate_multivariate(
                 flag_key,
                 &tenant_id.to_string(),
-                &self.collect_properties(tenant_id, tenant_properties),
+                &self.collect_properties(tenant_id),
             );
             match &res {
                 Ok(value) => {
@@ -276,7 +257,6 @@ impl FeatureResolver {
         &self,
         flag_key: &str,
         tenant_id: TenantId,
-        tenant_properties: &HashMap<String, PostHogFlagFilterPropertyValue>,
     ) -> Result<(), PostHogEvaluationError> {
         let force_overrides = self.force_overrides_for_testing.load();
         if let Some(value) = force_overrides.get(flag_key) {
@@ -291,7 +271,7 @@ impl FeatureResolver {
             let res = inner.feature_store().evaluate_boolean(
                 flag_key,
                 &tenant_id.to_string(),
-                &self.collect_properties(tenant_id, tenant_properties),
+                &self.collect_properties(tenant_id),
             );
             match &res {
                 Ok(()) => {
@@ -337,78 +317,3 @@ impl FeatureResolver {
             .store(Arc::new(force_overrides));
     }
 }
-
-struct PerTenantProperties {
-    pub remote_size_mb: Option<f64>,
-}
-
-impl PerTenantProperties {
-    pub fn into_posthog_properties(self) -> HashMap<String, PostHogFlagFilterPropertyValue> {
-        let mut properties = HashMap::new();
-        if let Some(remote_size_mb) = self.remote_size_mb {
-            properties.insert(
-                "tenant_remote_size_mb".to_string(),
-                PostHogFlagFilterPropertyValue::Number(remote_size_mb),
-            );
-        }
-        properties
-    }
-}
-
-#[derive(Clone)]
-pub struct TenantFeatureResolver {
-    inner: FeatureResolver,
-    tenant_id: TenantId,
-    cached_tenant_properties: Arc<ArcSwap<HashMap<String, PostHogFlagFilterPropertyValue>>>,
-}
-
-impl TenantFeatureResolver {
-    pub fn new(inner: FeatureResolver, tenant_id: TenantId) -> Self {
-        Self {
-            inner,
-            tenant_id,
-            cached_tenant_properties: Arc::new(ArcSwap::new(Arc::new(HashMap::new()))),
-        }
-    }
-
-    pub fn evaluate_multivariate(&self, flag_key: &str) -> Result<String, PostHogEvaluationError> {
-        self.inner.evaluate_multivariate(
-            flag_key,
-            self.tenant_id,
-            &self.cached_tenant_properties.load(),
-        )
-    }
-
-    pub fn evaluate_boolean(&self, flag_key: &str) -> Result<(), PostHogEvaluationError> {
-        self.inner.evaluate_boolean(
-            flag_key,
-            self.tenant_id,
-            &self.cached_tenant_properties.load(),
-        )
-    }
-
-    pub fn collect_properties(&self) -> HashMap<String, PostHogFlagFilterPropertyValue> {
-        self.inner
-            .collect_properties(self.tenant_id, &self.cached_tenant_properties.load())
-    }
-
-    pub fn is_feature_flag_boolean(&self, flag_key: &str) -> Result<bool, PostHogEvaluationError> {
-        self.inner.is_feature_flag_boolean(flag_key)
-    }
-
-    pub fn update_cached_tenant_properties(&self, tenant_shard: &TenantShard) {
-        let mut remote_size_mb = None;
-        for timeline in tenant_shard.list_timelines() {
-            let size = timeline.metrics.resident_physical_size_get();
-            if size == 0 {
-                remote_size_mb = None;
-            }
-            if let Some(ref mut remote_size_mb) = remote_size_mb {
-                *remote_size_mb += size as f64 / 1024.0 / 1024.0;
-            }
-        }
-        self.cached_tenant_properties.store(Arc::new(
-            PerTenantProperties { remote_size_mb }.into_posthog_properties(),
-        ));
-    }
-}

pageserver/src/http/routes.rs

@@ -2438,7 +2438,6 @@ async fn timeline_offload_handler(
             .map_err(|e| {
                 match e {
                     OffloadError::Cancelled => ApiError::ResourceUnavailable("Timeline shutting down".into()),
-                    OffloadError::AlreadyInProgress => ApiError::Conflict("Timeline already being offloaded or deleted".into()),
                     _ => ApiError::InternalServerError(anyhow!(e))
                 }
             })?;
@@ -3698,25 +3697,23 @@ async fn tenant_evaluate_feature_flag(
         let tenant = state
             .tenant_manager
             .get_attached_tenant_shard(tenant_shard_id)?;
-        // TODO: the properties we get here might be stale right after it is collected. But such races are rare (updated every 10s) 
-        // and we don't need to worry about it for now.
-        let properties = tenant.feature_resolver.collect_properties();
+        let properties = tenant.feature_resolver.collect_properties(tenant_shard_id.tenant_id);
         if as_type.as_deref() == Some("boolean") {
-            let result = tenant.feature_resolver.evaluate_boolean(&flag);
+            let result = tenant.feature_resolver.evaluate_boolean(&flag, tenant_shard_id.tenant_id);
             let result = result.map(|_| true).map_err(|e| e.to_string());
             json_response(StatusCode::OK, json!({ "result": result, "properties": properties }))
         } else if as_type.as_deref() == Some("multivariate") {
-            let result = tenant.feature_resolver.evaluate_multivariate(&flag).map_err(|e| e.to_string());
+            let result = tenant.feature_resolver.evaluate_multivariate(&flag, tenant_shard_id.tenant_id).map_err(|e| e.to_string());
             json_response(StatusCode::OK, json!({ "result": result, "properties": properties }))
         } else {
             // Auto infer the type of the feature flag.
             let is_boolean = tenant.feature_resolver.is_feature_flag_boolean(&flag).map_err(|e| ApiError::InternalServerError(anyhow::anyhow!("{e}")))?;
             if is_boolean {
-                let result = tenant.feature_resolver.evaluate_boolean(&flag);
+                let result = tenant.feature_resolver.evaluate_boolean(&flag, tenant_shard_id.tenant_id);
                 let result = result.map(|_| true).map_err(|e| e.to_string());
                 json_response(StatusCode::OK, json!({ "result": result, "properties": properties }))
             } else {
-                let result = tenant.feature_resolver.evaluate_multivariate(&flag).map_err(|e| e.to_string());
+                let result = tenant.feature_resolver.evaluate_multivariate(&flag, tenant_shard_id.tenant_id).map_err(|e| e.to_string());
                 json_response(StatusCode::OK, json!({ "result": result, "properties": properties }))
             }
         }

pageserver/src/tenant.rs

@@ -86,7 +86,7 @@ use crate::context;
 use crate::context::RequestContextBuilder;
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::deletion_queue::{DeletionQueueClient, DeletionQueueError};
-use crate::feature_resolver::{FeatureResolver, TenantFeatureResolver};
+use crate::feature_resolver::FeatureResolver;
 use crate::l0_flush::L0FlushGlobalState;
 use crate::metrics::{
     BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN, CONCURRENT_INITDBS,
@@ -386,7 +386,7 @@ pub struct TenantShard {
 
     l0_flush_global_state: L0FlushGlobalState,
 
-    pub(crate) feature_resolver: TenantFeatureResolver,
+    pub(crate) feature_resolver: FeatureResolver,
 }
 impl std::fmt::Debug for TenantShard {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
@@ -3263,7 +3263,7 @@ impl TenantShard {
                 };
                 let gc_compaction_strategy = self
                     .feature_resolver
-                    .evaluate_multivariate("gc-comapction-strategy")
+                    .evaluate_multivariate("gc-comapction-strategy", self.tenant_shard_id.tenant_id)
                     .ok();
                 let span = if let Some(gc_compaction_strategy) = gc_compaction_strategy {
                     info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id, strategy = %gc_compaction_strategy)
@@ -3285,7 +3285,6 @@ impl TenantShard {
                     .or_else(|err| match err {
                         // Ignore this, we likely raced with unarchival.
                         OffloadError::NotArchived => Ok(()),
-                        OffloadError::AlreadyInProgress => Ok(()),
                         err => Err(err),
                     })?;
             }
@@ -3409,9 +3408,6 @@ impl TenantShard {
         if let Some(ref walredo_mgr) = self.walredo_mgr {
             walredo_mgr.maybe_quiesce(WALREDO_IDLE_TIMEOUT);
         }
-
-        // Update the feature resolver with the latest tenant-spcific data.
-        self.feature_resolver.update_cached_tenant_properties(self);
     }
 
     pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
@@ -4494,10 +4490,7 @@ impl TenantShard {
             gc_block: Default::default(),
             l0_flush_global_state,
             basebackup_cache,
-            feature_resolver: TenantFeatureResolver::new(
-                feature_resolver,
-                tenant_shard_id.tenant_id,
-            ),
+            feature_resolver,
         }
     }
 

pageserver/src/tenant/storage_layer/batch_split_writer.rs

@@ -182,7 +182,7 @@ impl BatchLayerWriter {
 /// An image writer that takes images and produces multiple image layers.
 #[must_use]
 pub struct SplitImageLayerWriter<'a> {
-    inner: Option<ImageLayerWriter>,
+    inner: ImageLayerWriter,
     target_layer_size: u64,
     lsn: Lsn,
     conf: &'static PageServerConf,
@@ -196,7 +196,7 @@ pub struct SplitImageLayerWriter<'a> {
 
 impl<'a> SplitImageLayerWriter<'a> {
     #[allow(clippy::too_many_arguments)]
-    pub fn new(
+    pub async fn new(
         conf: &'static PageServerConf,
         timeline_id: TimelineId,
         tenant_shard_id: TenantShardId,
@@ -205,10 +205,22 @@ impl<'a> SplitImageLayerWriter<'a> {
         target_layer_size: u64,
         gate: &'a utils::sync::gate::Gate,
         cancel: CancellationToken,
-    ) -> Self {
-        Self {
+        ctx: &RequestContext,
+    ) -> anyhow::Result<Self> {
+        Ok(Self {
             target_layer_size,
-            inner: None,
+            // XXX make this lazy like in SplitDeltaLayerWriter?
+            inner: ImageLayerWriter::new(
+                conf,
+                timeline_id,
+                tenant_shard_id,
+                &(start_key..Key::MAX),
+                lsn,
+                gate,
+                cancel.clone(),
+                ctx,
+            )
+            .await?,
             conf,
             timeline_id,
             tenant_shard_id,
@@ -217,7 +229,7 @@ impl<'a> SplitImageLayerWriter<'a> {
             start_key,
             gate,
             cancel,
-        }
+        })
     }
 
     pub async fn put_image(
@@ -226,31 +238,12 @@ impl<'a> SplitImageLayerWriter<'a> {
         img: Bytes,
         ctx: &RequestContext,
     ) -> Result<(), PutError> {
-        if self.inner.is_none() {
-            self.inner = Some(
-                ImageLayerWriter::new(
-                    self.conf,
-                    self.timeline_id,
-                    self.tenant_shard_id,
-                    &(self.start_key..Key::MAX),
-                    self.lsn,
-                    self.gate,
-                    self.cancel.clone(),
-                    ctx,
-                )
-                .await
-                .map_err(PutError::Other)?,
-            );
-        }
-
-        let inner = self.inner.as_mut().unwrap();
-
         // The current estimation is an upper bound of the space that the key/image could take
         // because we did not consider compression in this estimation. The resulting image layer
         // could be smaller than the target size.
         let addition_size_estimation = KEY_SIZE as u64 + img.len() as u64;
-        if inner.num_keys() >= 1
-            && inner.estimated_size() + addition_size_estimation >= self.target_layer_size
+        if self.inner.num_keys() >= 1
+            && self.inner.estimated_size() + addition_size_estimation >= self.target_layer_size
         {
             let next_image_writer = ImageLayerWriter::new(
                 self.conf,
@@ -264,7 +257,7 @@ impl<'a> SplitImageLayerWriter<'a> {
             )
             .await
             .map_err(PutError::Other)?;
-            let prev_image_writer = std::mem::replace(inner, next_image_writer);
+            let prev_image_writer = std::mem::replace(&mut self.inner, next_image_writer);
             self.batches.add_unfinished_image_writer(
                 prev_image_writer,
                 self.start_key..key,
@@ -272,7 +265,7 @@ impl<'a> SplitImageLayerWriter<'a> {
             );
             self.start_key = key;
         }
-        inner.put_image(key, img, ctx).await
+        self.inner.put_image(key, img, ctx).await
     }
 
     pub(crate) async fn finish_with_discard_fn<D, F>(
@@ -289,10 +282,8 @@ impl<'a> SplitImageLayerWriter<'a> {
         let Self {
             mut batches, inner, ..
         } = self;
-        if let Some(inner) = inner {
-            if inner.num_keys() != 0 {
-                batches.add_unfinished_image_writer(inner, self.start_key..end_key, self.lsn);
-            }
+        if inner.num_keys() != 0 {
+            batches.add_unfinished_image_writer(inner, self.start_key..end_key, self.lsn);
         }
         batches.finish_with_discard_fn(tline, ctx, discard_fn).await
     }
@@ -507,7 +498,10 @@ mod tests {
             4 * 1024 * 1024,
             &tline.gate,
             tline.cancel.clone(),
-        );
+            &ctx,
+        )
+        .await
+        .unwrap();
 
         let mut delta_writer = SplitDeltaLayerWriter::new(
             tenant.conf,
@@ -583,7 +577,10 @@ mod tests {
             4 * 1024 * 1024,
             &tline.gate,
             tline.cancel.clone(),
-        );
+            &ctx,
+        )
+        .await
+        .unwrap();
         let mut delta_writer = SplitDeltaLayerWriter::new(
             tenant.conf,
             tline.timeline_id,
@@ -679,7 +676,10 @@ mod tests {
             4 * 1024,
             &tline.gate,
             tline.cancel.clone(),
-        );
+            &ctx,
+        )
+        .await
+        .unwrap();
 
         let mut delta_writer = SplitDeltaLayerWriter::new(
             tenant.conf,

pageserver/src/tenant/timeline.rs

@@ -106,7 +106,7 @@ use crate::context::{
     DownloadBehavior, PerfInstrumentFutureExt, RequestContext, RequestContextBuilder,
 };
 use crate::disk_usage_eviction_task::{DiskUsageEvictionInfo, EvictionCandidate, finite_f32};
-use crate::feature_resolver::TenantFeatureResolver;
+use crate::feature_resolver::FeatureResolver;
 use crate::keyspace::{KeyPartitioning, KeySpace};
 use crate::l0_flush::{self, L0FlushGlobalState};
 use crate::metrics::{
@@ -202,7 +202,7 @@ pub struct TimelineResources {
     pub l0_compaction_trigger: Arc<Notify>,
     pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
     pub basebackup_cache: Arc<BasebackupCache>,
-    pub feature_resolver: TenantFeatureResolver,
+    pub feature_resolver: FeatureResolver,
 }
 
 pub struct Timeline {
@@ -450,7 +450,7 @@ pub struct Timeline {
     /// A channel to send async requests to prepare a basebackup for the basebackup cache.
     basebackup_cache: Arc<BasebackupCache>,
 
-    feature_resolver: TenantFeatureResolver,
+    feature_resolver: FeatureResolver,
 }
 
 pub(crate) enum PreviousHeatmap {
@@ -5308,7 +5308,6 @@ impl Timeline {
         ctx: &RequestContext,
         img_range: Range<Key>,
         io_concurrency: IoConcurrency,
-        progress: Option<(usize, usize)>,
     ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
         let mut wrote_keys = false;
 
@@ -5385,15 +5384,11 @@ impl Timeline {
             }
         }
 
-        let progress_report = progress
-            .map(|(idx, total)| format!("({idx}/{total}) "))
-            .unwrap_or_default();
         if wrote_keys {
             // Normal path: we have written some data into the new image layer for this
             // partition, so flush it to disk.
             info!(
-                "{} produced image layer for rel {}",
-                progress_report,
+                "produced image layer for rel {}",
                 ImageLayerName {
                     key_range: img_range.clone(),
                     lsn
@@ -5403,12 +5398,7 @@ impl Timeline {
                 unfinished_image_layer: image_layer_writer,
             })
         } else {
-            tracing::debug!(
-                "{} no data in range {}-{}",
-                progress_report,
-                img_range.start,
-                img_range.end
-            );
+            tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
             Ok(ImageLayerCreationOutcome::Empty)
         }
     }
@@ -5643,8 +5633,7 @@ impl Timeline {
             }
         }
 
-        let total = partition_parts.len();
-        for (idx, partition) in partition_parts.iter().enumerate() {
+        for partition in partition_parts.iter() {
             if self.cancel.is_cancelled() {
                 return Err(CreateImageLayersError::Cancelled);
             }
@@ -5729,7 +5718,6 @@ impl Timeline {
                     ctx,
                     img_range.clone(),
                     io_concurrency,
-                    Some((idx, total)),
                 )
                 .await?
             } else {

pageserver/src/tenant/timeline/compaction.rs

@@ -9,7 +9,7 @@ use std::ops::{Deref, Range};
 use std::sync::Arc;
 use std::time::{Duration, Instant};
 
-use super::layer_manager::LayerManagerLockHolder;
+use super::layer_manager::{LayerManagerLockHolder, LayerManagerReadGuard};
 use super::{
     CompactFlags, CompactOptions, CompactionError, CreateImageLayersError, DurationRecorder,
     GetVectoredError, ImageLayerCreationMode, LastImageLayerCreationStatus, RecordedDuration,
@@ -101,11 +101,7 @@ pub enum GcCompactionQueueItem {
         /// Whether the compaction is triggered automatically (determines whether we need to update L2 LSN)
         auto: bool,
     },
-    SubCompactionJob {
-        i: usize,
-        total: usize,
-        options: CompactOptions,
-    },
+    SubCompactionJob(CompactOptions),
     Notify(GcCompactionJobId, Option<Lsn>),
 }
 
@@ -167,7 +163,7 @@ impl GcCompactionQueueItem {
                 running,
                 job_id: id.0,
             }),
-            GcCompactionQueueItem::SubCompactionJob { options, .. } => Some(CompactInfoResponse {
+            GcCompactionQueueItem::SubCompactionJob(options) => Some(CompactInfoResponse {
                 compact_key_range: options.compact_key_range,
                 compact_lsn_range: options.compact_lsn_range,
                 sub_compaction: options.sub_compaction,
@@ -493,7 +489,7 @@ impl GcCompactionQueue {
                 .map(|job| job.compact_lsn_range.end)
                 .max()
                 .unwrap();
-            for (i, job) in jobs.into_iter().enumerate() {
+            for job in jobs {
                 // Unfortunately we need to convert the `GcCompactJob` back to `CompactionOptions`
                 // until we do further refactors to allow directly call `compact_with_gc`.
                 let mut flags: EnumSet<CompactFlags> = EnumSet::default();
@@ -511,11 +507,7 @@ impl GcCompactionQueue {
                     compact_lsn_range: Some(job.compact_lsn_range.into()),
                     sub_compaction_max_job_size_mb: None,
                 };
-                pending_tasks.push(GcCompactionQueueItem::SubCompactionJob {
-                    options,
-                    i,
-                    total: jobs_len,
-                });
+                pending_tasks.push(GcCompactionQueueItem::SubCompactionJob(options));
             }
 
             if !auto {
@@ -659,7 +651,7 @@ impl GcCompactionQueue {
                     }
                 }
             }
-            GcCompactionQueueItem::SubCompactionJob { options, i, total } => {
+            GcCompactionQueueItem::SubCompactionJob(options) => {
                 // TODO: error handling, clear the queue if any task fails?
                 let _gc_guard = match gc_block.start().await {
                     Ok(guard) => guard,
@@ -671,7 +663,6 @@ impl GcCompactionQueue {
                         )));
                     }
                 };
-                info!("running gc-compaction subcompaction job {}/{}", i, total);
                 let res = timeline.compact_with_options(cancel, options, ctx).await;
                 let compaction_result = match res {
                     Ok(res) => res,
@@ -1319,7 +1310,7 @@ impl Timeline {
             || cfg!(feature = "testing")
             || self
                 .feature_resolver
-                .evaluate_boolean("image-compaction-boundary")
+                .evaluate_boolean("image-compaction-boundary", self.tenant_shard_id.tenant_id)
                 .is_ok()
         {
             let last_repartition_lsn = self.partitioning.read().1;
@@ -1600,15 +1591,13 @@ impl Timeline {
         let started = Instant::now();
 
         let mut replace_image_layers = Vec::new();
-        let total = layers_to_rewrite.len();
 
-        for (i, layer) in layers_to_rewrite.into_iter().enumerate() {
+        for layer in layers_to_rewrite {
             if self.cancel.is_cancelled() {
                 return Err(CompactionError::ShuttingDown);
             }
 
-            info!(layer=%layer, "rewriting layer after shard split: {}/{}", i, total);
-
+            info!(layer=%layer, "rewriting layer after shard split");
             let mut image_layer_writer = ImageLayerWriter::new(
                 self.conf,
                 self.timeline_id,
@@ -1790,14 +1779,20 @@ impl Timeline {
         } = {
             let phase1_span = info_span!("compact_level0_phase1");
             let ctx = ctx.attached_child();
-            let stats = CompactLevel0Phase1StatsBuilder {
+            let mut stats = CompactLevel0Phase1StatsBuilder {
                 version: Some(2),
                 tenant_id: Some(self.tenant_shard_id),
                 timeline_id: Some(self.timeline_id),
                 ..Default::default()
             };
 
+            let begin = tokio::time::Instant::now();
+            let phase1_layers_locked = self.layers.read(LayerManagerLockHolder::Compaction).await;
+            let now = tokio::time::Instant::now();
+            stats.read_lock_acquisition_micros =
+                DurationRecorder::Recorded(RecordedDuration(now - begin), now);
             self.compact_level0_phase1(
+                phase1_layers_locked,
                 stats,
                 target_file_size,
                 force_compaction_ignore_threshold,
@@ -1818,19 +1813,16 @@ impl Timeline {
     }
 
     /// Level0 files first phase of compaction, explained in the [`Self::compact_legacy`] comment.
-    async fn compact_level0_phase1(
-        self: &Arc<Self>,
+    async fn compact_level0_phase1<'a>(
+        self: &'a Arc<Self>,
+        guard: LayerManagerReadGuard<'a>,
         mut stats: CompactLevel0Phase1StatsBuilder,
         target_file_size: u64,
         force_compaction_ignore_threshold: bool,
         ctx: &RequestContext,
     ) -> Result<CompactLevel0Phase1Result, CompactionError> {
-        let begin = tokio::time::Instant::now();
-        let guard = self.layers.read(LayerManagerLockHolder::Compaction).await;
-        let now = tokio::time::Instant::now();
-        stats.read_lock_acquisition_micros =
-            DurationRecorder::Recorded(RecordedDuration(now - begin), now);
-
+        stats.read_lock_held_spawn_blocking_startup_micros =
+            stats.read_lock_acquisition_micros.till_now(); // set by caller
         let layers = guard.layer_map()?;
         let level0_deltas = layers.level0_deltas();
         stats.level0_deltas_count = Some(level0_deltas.len());
@@ -1865,12 +1857,6 @@ impl Timeline {
             .map(|x| guard.get_from_desc(x))
             .collect::<Vec<_>>();
 
-        drop_layer_manager_rlock(guard);
-
-        // The is the last LSN that we have seen for L0 compaction in the timeline. This LSN might be updated
-        // by the time we finish the compaction. So we need to get it here.
-        let l0_last_record_lsn = self.get_last_record_lsn();
-
         // Gather the files to compact in this iteration.
         //
         // Start with the oldest Level 0 delta file, and collect any other
@@ -1958,7 +1944,9 @@ impl Timeline {
         // we don't accidentally use it later in the function.
         drop(level0_deltas);
 
-        stats.compaction_prerequisites_micros = stats.read_lock_acquisition_micros.till_now();
+        stats.read_lock_held_prerequisites_micros = stats
+            .read_lock_held_spawn_blocking_startup_micros
+            .till_now();
 
         // TODO: replace with streaming k-merge
         let all_keys = {
@@ -1980,7 +1968,7 @@ impl Timeline {
             all_keys
         };
 
-        stats.read_lock_held_key_sort_micros = stats.compaction_prerequisites_micros.till_now();
+        stats.read_lock_held_key_sort_micros = stats.read_lock_held_prerequisites_micros.till_now();
 
         // Determine N largest holes where N is number of compacted layers. The vec is sorted by key range start.
         //
@@ -2014,6 +2002,7 @@ impl Timeline {
                 }
             }
             let max_holes = deltas_to_compact.len();
+            let last_record_lsn = self.get_last_record_lsn();
             let min_hole_range = (target_file_size / page_cache::PAGE_SZ as u64) as i128;
             let min_hole_coverage_size = 3; // TODO: something more flexible?
             // min-heap (reserve space for one more element added before eviction)
@@ -2032,12 +2021,8 @@ impl Timeline {
                         // has not so much sense, because largest holes will corresponds field1/field2 changes.
                         // But we are mostly interested to eliminate holes which cause generation of excessive image layers.
                         // That is why it is better to measure size of hole as number of covering image layers.
-                        let coverage_size = {
-                            // TODO: optimize this with copy-on-write layer map.
-                            let guard = self.layers.read(LayerManagerLockHolder::Compaction).await;
-                            let layers = guard.layer_map()?;
-                            layers.image_coverage(&key_range, l0_last_record_lsn).len()
-                        };
+                        let coverage_size =
+                            layers.image_coverage(&key_range, last_record_lsn).len();
                         if coverage_size >= min_hole_coverage_size {
                             heap.push(Hole {
                                 key_range,
@@ -2056,6 +2041,7 @@ impl Timeline {
             holes
         };
         stats.read_lock_held_compute_holes_micros = stats.read_lock_held_key_sort_micros.till_now();
+        drop_layer_manager_rlock(guard);
 
         if self.cancel.is_cancelled() {
             return Err(CompactionError::ShuttingDown);
@@ -2396,8 +2382,9 @@ struct CompactLevel0Phase1StatsBuilder {
     tenant_id: Option<TenantShardId>,
     timeline_id: Option<TimelineId>,
     read_lock_acquisition_micros: DurationRecorder,
+    read_lock_held_spawn_blocking_startup_micros: DurationRecorder,
     read_lock_held_key_sort_micros: DurationRecorder,
-    compaction_prerequisites_micros: DurationRecorder,
+    read_lock_held_prerequisites_micros: DurationRecorder,
     read_lock_held_compute_holes_micros: DurationRecorder,
     read_lock_drop_micros: DurationRecorder,
     write_layer_files_micros: DurationRecorder,
@@ -2412,8 +2399,9 @@ struct CompactLevel0Phase1Stats {
     tenant_id: TenantShardId,
     timeline_id: TimelineId,
     read_lock_acquisition_micros: RecordedDuration,
+    read_lock_held_spawn_blocking_startup_micros: RecordedDuration,
     read_lock_held_key_sort_micros: RecordedDuration,
-    compaction_prerequisites_micros: RecordedDuration,
+    read_lock_held_prerequisites_micros: RecordedDuration,
     read_lock_held_compute_holes_micros: RecordedDuration,
     read_lock_drop_micros: RecordedDuration,
     write_layer_files_micros: RecordedDuration,
@@ -2438,12 +2426,16 @@ impl TryFrom<CompactLevel0Phase1StatsBuilder> for CompactLevel0Phase1Stats {
                 .read_lock_acquisition_micros
                 .into_recorded()
                 .ok_or_else(|| anyhow!("read_lock_acquisition_micros not set"))?,
+            read_lock_held_spawn_blocking_startup_micros: value
+                .read_lock_held_spawn_blocking_startup_micros
+                .into_recorded()
+                .ok_or_else(|| anyhow!("read_lock_held_spawn_blocking_startup_micros not set"))?,
             read_lock_held_key_sort_micros: value
                 .read_lock_held_key_sort_micros
                 .into_recorded()
                 .ok_or_else(|| anyhow!("read_lock_held_key_sort_micros not set"))?,
-            compaction_prerequisites_micros: value
-                .compaction_prerequisites_micros
+            read_lock_held_prerequisites_micros: value
+                .read_lock_held_prerequisites_micros
                 .into_recorded()
                 .ok_or_else(|| anyhow!("read_lock_held_prerequisites_micros not set"))?,
             read_lock_held_compute_holes_micros: value
@@ -3511,16 +3503,22 @@ impl Timeline {
         // Only create image layers when there is no ancestor branches. TODO: create covering image layer
         // when some condition meet.
         let mut image_layer_writer = if !has_data_below {
-            Some(SplitImageLayerWriter::new(
-                self.conf,
-                self.timeline_id,
-                self.tenant_shard_id,
-                job_desc.compaction_key_range.start,
-                lowest_retain_lsn,
-                self.get_compaction_target_size(),
-                &self.gate,
-                self.cancel.clone(),
-            ))
+            Some(
+                SplitImageLayerWriter::new(
+                    self.conf,
+                    self.timeline_id,
+                    self.tenant_shard_id,
+                    job_desc.compaction_key_range.start,
+                    lowest_retain_lsn,
+                    self.get_compaction_target_size(),
+                    &self.gate,
+                    self.cancel.clone(),
+                    ctx,
+                )
+                .await
+                .context("failed to create image layer writer")
+                .map_err(CompactionError::Other)?,
+            )
         } else {
             None
         };
@@ -4354,7 +4352,6 @@ impl TimelineAdaptor {
                 ctx,
                 key_range.clone(),
                 IoConcurrency::sequential(),
-                None,
             )
             .await?;
 

pageserver/src/tenant/timeline/offload.rs

@@ -19,8 +19,6 @@ pub(crate) enum OffloadError {
     NotArchived,
     #[error(transparent)]
     RemoteStorage(anyhow::Error),
-    #[error("Offload or deletion already in progress")]
-    AlreadyInProgress,
     #[error("Unexpected offload error: {0}")]
     Other(anyhow::Error),
 }
@@ -46,26 +44,20 @@ pub(crate) async fn offload_timeline(
         timeline.timeline_id,
         TimelineDeleteGuardKind::Offload,
     );
-    let (timeline, guard) = match delete_guard_res {
-        Ok(timeline_and_guard) => timeline_and_guard,
-        Err(DeleteTimelineError::HasChildren(children)) => {
-            let is_archived = timeline.is_archived();
-            if is_archived == Some(true) {
-                tracing::error!("timeline is archived but has non-archived children: {children:?}");
-                return Err(OffloadError::NotArchived);
-            }
-            tracing::info!(
-                ?is_archived,
-                "timeline is not archived and has unarchived children"
-            );
+    if let Err(DeleteTimelineError::HasChildren(children)) = delete_guard_res {
+        let is_archived = timeline.is_archived();
+        if is_archived == Some(true) {
+            tracing::error!("timeline is archived but has non-archived children: {children:?}");
             return Err(OffloadError::NotArchived);
         }
-        Err(DeleteTimelineError::AlreadyInProgress(_)) => {
-            tracing::info!("timeline offload or deletion already in progress");
-            return Err(OffloadError::AlreadyInProgress);
-        }
-        Err(e) => return Err(OffloadError::Other(anyhow::anyhow!(e))),
+        tracing::info!(
+            ?is_archived,
+            "timeline is not archived and has unarchived children"
+        );
+        return Err(OffloadError::NotArchived);
     };
+    let (timeline, guard) =
+        delete_guard_res.map_err(|e| OffloadError::Other(anyhow::anyhow!(e)))?;
 
     let TimelineOrOffloaded::Timeline(timeline) = timeline else {
         tracing::error!("timeline already offloaded, but given timeline object");

proxy/src/binary/pg_sni_router.rs

@@ -4,7 +4,6 @@
 //! This allows connecting to pods/services running in the same Kubernetes cluster from
 //! the outside. Similar to an ingress controller for HTTPS.
 
-use std::io;
 use std::net::SocketAddr;
 use std::path::Path;
 use std::sync::Arc;
@@ -230,6 +229,7 @@ pub(super) async fn task_main(
                     .set_nodelay(true)
                     .context("failed to set socket option")?;
 
+                info!(%peer_addr, "serving");
                 let ctx = RequestContext::new(
                     session_id,
                     ConnectionInfo {
@@ -241,14 +241,6 @@ pub(super) async fn task_main(
                 handle_client(ctx, dest_suffix, tls_config, compute_tls_config, socket).await
             }
             .unwrap_or_else(|e| {
-                if let Some(FirstMessage(io_error)) = e.downcast_ref() {
-                    // this is noisy. if we get EOF on the very first message that's likely
-                    // just NLB doing a healthcheck.
-                    if io_error.kind() == io::ErrorKind::UnexpectedEof {
-                        return;
-                    }
-                }
-
                 // Acknowledge that the task has finished with an error.
                 error!("per-client task finished with an error: {e:#}");
             })
@@ -265,19 +257,12 @@ pub(super) async fn task_main(
     Ok(())
 }
 
-#[derive(Debug, thiserror::Error)]
-#[error(transparent)]
-struct FirstMessage(io::Error);
-
 async fn ssl_handshake<S: AsyncRead + AsyncWrite + Unpin>(
     ctx: &RequestContext,
     raw_stream: S,
     tls_config: Arc<rustls::ServerConfig>,
 ) -> anyhow::Result<TlsStream<S>> {
-    let (mut stream, msg) = PqStream::parse_startup(Stream::from_raw(raw_stream))
-        .await
-        .map_err(FirstMessage)?;
-
+    let (mut stream, msg) = PqStream::parse_startup(Stream::from_raw(raw_stream)).await?;
     match msg {
         FeStartupPacket::SslRequest { direct: None } => {
             let raw = stream.accept_tls().await?;

proxy/src/control_plane/client/mod.rs

@@ -213,12 +213,7 @@ impl<K: Hash + Eq + Clone> ApiLocks<K> {
         self.metrics
             .semaphore_acquire_seconds
             .observe(now.elapsed().as_secs_f64());
-
-        if permit.is_ok() {
-            debug!(elapsed = ?now.elapsed(), "acquired permit");
-        } else {
-            debug!(elapsed = ?now.elapsed(), "timed out acquiring permit");
-        }
+        debug!("acquired permit {:?}", now.elapsed().as_secs_f64());
         Ok(WakeComputePermit { permit: permit? })
     }
 

proxy/src/error.rs

@@ -78,6 +78,16 @@ pub(crate) trait ReportableError: fmt::Display + Send + 'static {
     fn get_error_kind(&self) -> ErrorKind;
 }
 
+impl ReportableError for postgres_client::error::Error {
+    fn get_error_kind(&self) -> ErrorKind {
+        if self.as_db_error().is_some() {
+            ErrorKind::Postgres
+        } else {
+            ErrorKind::Compute
+        }
+    }
+}
+
 /// Flattens `Result<Result<T>>` into `Result<T>`.
 pub fn flatten_err<T>(r: Result<anyhow::Result<T>, JoinError>) -> anyhow::Result<T> {
     r.context("join error").and_then(|x| x)

proxy/src/logging.rs

@@ -52,7 +52,7 @@ pub async fn init() -> anyhow::Result<LoggingGuard> {
             StderrWriter {
                 stderr: std::io::stderr(),
             },
-            &["conn_id", "ep", "query_id", "request_id", "session_id"],
+            &["request_id", "session_id", "conn_id"],
         ))
     } else {
         None

proxy/src/redis/connection_with_credentials_provider.rs

@@ -1,3 +1,4 @@
+use std::convert::Infallible;
 use std::sync::{Arc, atomic::AtomicBool, atomic::Ordering};
 use std::time::Duration;
 
@@ -5,7 +6,7 @@ use futures::FutureExt;
 use redis::aio::{ConnectionLike, MultiplexedConnection};
 use redis::{ConnectionInfo, IntoConnectionInfo, RedisConnectionInfo, RedisResult};
 use tokio::task::JoinHandle;
-use tracing::{debug, error, info, warn};
+use tracing::{error, info, warn};
 
 use super::elasticache::CredentialsProvider;
 
@@ -31,7 +32,7 @@ pub struct ConnectionWithCredentialsProvider {
     credentials: Credentials,
     // TODO: with more load on the connection, we should consider using a connection pool
     con: Option<MultiplexedConnection>,
-    refresh_token_task: Option<JoinHandle<()>>,
+    refresh_token_task: Option<JoinHandle<Infallible>>,
     mutex: tokio::sync::Mutex<()>,
     credentials_refreshed: Arc<AtomicBool>,
 }
@@ -121,15 +122,11 @@ impl ConnectionWithCredentialsProvider {
             let credentials_provider = credentials_provider.clone();
             let con2 = con.clone();
             let credentials_refreshed = self.credentials_refreshed.clone();
-            let f = tokio::spawn(async move {
-                let result = Self::keep_connection(con2, credentials_provider).await;
-                if let Err(e) = result {
-                    credentials_refreshed.store(false, Ordering::Release);
-                    debug!("keep_connection failed: {e}");
-                } else {
-                    credentials_refreshed.store(true, Ordering::Release);
-                }
-            });
+            let f = tokio::spawn(Self::keep_connection(
+                con2,
+                credentials_provider,
+                credentials_refreshed,
+            ));
             self.refresh_token_task = Some(f);
         }
         match Self::ping(&mut con).await {
@@ -165,6 +162,7 @@ impl ConnectionWithCredentialsProvider {
 
     async fn get_client(&self) -> anyhow::Result<redis::Client> {
         let client = redis::Client::open(self.get_connection_info().await?)?;
+        self.credentials_refreshed.store(true, Ordering::Relaxed);
         Ok(client)
     }
 
@@ -180,16 +178,19 @@ impl ConnectionWithCredentialsProvider {
     async fn keep_connection(
         mut con: MultiplexedConnection,
         credentials_provider: Arc<CredentialsProvider>,
-    ) -> anyhow::Result<()> {
+        credentials_refreshed: Arc<AtomicBool>,
+    ) -> Infallible {
         loop {
             // The connection lives for 12h, for the sanity check we refresh it every hour.
             tokio::time::sleep(Duration::from_secs(60 * 60)).await;
             match Self::refresh_token(&mut con, credentials_provider.clone()).await {
                 Ok(()) => {
                     info!("Token refreshed");
+                    credentials_refreshed.store(true, Ordering::Relaxed);
                 }
                 Err(e) => {
                     error!("Error during token refresh: {e:?}");
+                    credentials_refreshed.store(false, Ordering::Relaxed);
                 }
             }
         }

proxy/src/serverless/backend.rs

@@ -404,15 +404,7 @@ impl ReportableError for HttpConnError {
     fn get_error_kind(&self) -> ErrorKind {
         match self {
             HttpConnError::ConnectionClosedAbruptly(_) => ErrorKind::Compute,
-            HttpConnError::PostgresConnectionError(p) => {
-                if p.as_db_error().is_some() {
-                    // postgres rejected the connection
-                    ErrorKind::Postgres
-                } else {
-                    // couldn't even reach postgres
-                    ErrorKind::Compute
-                }
-            }
+            HttpConnError::PostgresConnectionError(p) => p.get_error_kind(),
             HttpConnError::LocalProxyConnectionError(_) => ErrorKind::Compute,
             HttpConnError::ComputeCtl(_) => ErrorKind::Service,
             HttpConnError::JwtPayloadError(_) => ErrorKind::User,

proxy/src/serverless/json.rs

@@ -70,34 +70,6 @@ pub(crate) enum JsonConversionError {
     ParseJsonError(#[from] serde_json::Error),
     #[error("unbalanced array")]
     UnbalancedArray,
-    #[error("unbalanced quoted string")]
-    UnbalancedString,
-}
-
-enum OutputMode {
-    Array(Vec<Value>),
-    Object(Map<String, Value>),
-}
-
-impl OutputMode {
-    fn key(&mut self, key: &str) -> &mut Value {
-        match self {
-            OutputMode::Array(values) => push_entry(values, Value::Null),
-            OutputMode::Object(map) => map.entry(key.to_string()).or_insert(Value::Null),
-        }
-    }
-
-    fn finish(self) -> Value {
-        match self {
-            OutputMode::Array(values) => Value::Array(values),
-            OutputMode::Object(map) => Value::Object(map),
-        }
-    }
-}
-
-fn push_entry<T>(arr: &mut Vec<T>, t: T) -> &mut T {
-    arr.push(t);
-    arr.last_mut().expect("a value was just inserted")
 }
 
 //
@@ -105,277 +77,182 @@ fn push_entry<T>(arr: &mut Vec<T>, t: T) -> &mut T {
 //
 pub(crate) fn pg_text_row_to_json(
     row: &Row,
+    columns: &[Type],
     raw_output: bool,
     array_mode: bool,
 ) -> Result<Value, JsonConversionError> {
-    let mut entries = if array_mode {
-        OutputMode::Array(Vec::with_capacity(row.columns().len()))
+    let iter = row
+        .columns()
+        .iter()
+        .zip(columns)
+        .enumerate()
+        .map(|(i, (column, typ))| {
+            let name = column.name();
+            let pg_value = row.as_text(i).map_err(JsonConversionError::AsTextError)?;
+            let json_value = if raw_output {
+                match pg_value {
+                    Some(v) => Value::String(v.to_string()),
+                    None => Value::Null,
+                }
+            } else {
+                pg_text_to_json(pg_value, typ)?
+            };
+            Ok((name.to_string(), json_value))
+        });
+
+    if array_mode {
+        // drop keys and aggregate into array
+        let arr = iter
+            .map(|r| r.map(|(_key, val)| val))
+            .collect::<Result<Vec<Value>, JsonConversionError>>()?;
+        Ok(Value::Array(arr))
     } else {
-        OutputMode::Object(Map::with_capacity(row.columns().len()))
-    };
-
-    for (i, column) in row.columns().iter().enumerate() {
-        let pg_value = row.as_text(i).map_err(JsonConversionError::AsTextError)?;
-
-        let value = entries.key(column.name());
-
-        match pg_value {
-            Some(v) if raw_output => *value = Value::String(v.to_string()),
-            Some(v) => pg_text_to_json(value, v, column.type_())?,
-            None => *value = Value::Null,
-        }
+        let obj = iter.collect::<Result<Map<String, Value>, JsonConversionError>>()?;
+        Ok(Value::Object(obj))
     }
-
-    Ok(entries.finish())
 }
 
 //
 // Convert postgres text-encoded value to JSON value
 //
-fn pg_text_to_json(
-    output: &mut Value,
-    val: &str,
-    pg_type: &Type,
-) -> Result<(), JsonConversionError> {
-    if let Kind::Array(elem_type) = pg_type.kind() {
-        // todo: we should fetch this from postgres.
-        let delimiter = ',';
-
-        let mut array = vec![];
-        pg_array_parse(&mut array, val, elem_type, delimiter)?;
-        *output = Value::Array(array);
-        return Ok(());
-    }
-
-    match *pg_type {
-        Type::BOOL => *output = Value::Bool(val == "t"),
-        Type::INT2 | Type::INT4 => {
-            let val = val.parse::<i32>()?;
-            *output = Value::Number(serde_json::Number::from(val));
+fn pg_text_to_json(pg_value: Option<&str>, pg_type: &Type) -> Result<Value, JsonConversionError> {
+    if let Some(val) = pg_value {
+        if let Kind::Array(elem_type) = pg_type.kind() {
+            return pg_array_parse(val, elem_type);
         }
-        Type::FLOAT4 | Type::FLOAT8 => {
-            let fval = val.parse::<f64>()?;
-            let num = serde_json::Number::from_f64(fval);
-            if let Some(num) = num {
-                *output = Value::Number(num);
-            } else {
-                // Pass Nan, Inf, -Inf as strings
-                // JS JSON.stringify() does converts them to null, but we
-                // want to preserve them, so we pass them as strings
-                *output = Value::String(val.to_string());
+
+        match *pg_type {
+            Type::BOOL => Ok(Value::Bool(val == "t")),
+            Type::INT2 | Type::INT4 => {
+                let val = val.parse::<i32>()?;
+                Ok(Value::Number(serde_json::Number::from(val)))
+            }
+            Type::FLOAT4 | Type::FLOAT8 => {
+                let fval = val.parse::<f64>()?;
+                let num = serde_json::Number::from_f64(fval);
+                if let Some(num) = num {
+                    Ok(Value::Number(num))
+                } else {
+                    // Pass Nan, Inf, -Inf as strings
+                    // JS JSON.stringify() does converts them to null, but we
+                    // want to preserve them, so we pass them as strings
+                    Ok(Value::String(val.to_string()))
+                }
+            }
+            Type::JSON | Type::JSONB => Ok(serde_json::from_str(val)?),
+            _ => Ok(Value::String(val.to_string())),
+        }
+    } else {
+        Ok(Value::Null)
+    }
+}
+
+//
+// Parse postgres array into JSON array.
+//
+// This is a bit involved because we need to handle nested arrays and quoted
+// values. Unlike postgres we don't check that all nested arrays have the same
+// dimensions, we just return them as is.
+//
+fn pg_array_parse(pg_array: &str, elem_type: &Type) -> Result<Value, JsonConversionError> {
+    pg_array_parse_inner(pg_array, elem_type, false).map(|(v, _)| v)
+}
+
+fn pg_array_parse_inner(
+    pg_array: &str,
+    elem_type: &Type,
+    nested: bool,
+) -> Result<(Value, usize), JsonConversionError> {
+    let mut pg_array_chr = pg_array.char_indices();
+    let mut level = 0;
+    let mut quote = false;
+    let mut entries: Vec<Value> = Vec::new();
+    let mut entry = String::new();
+
+    // skip bounds decoration
+    if let Some('[') = pg_array.chars().next() {
+        for (_, c) in pg_array_chr.by_ref() {
+            if c == '=' {
+                break;
             }
         }
-        Type::JSON | Type::JSONB => *output = serde_json::from_str(val)?,
-        _ => *output = Value::String(val.to_string()),
     }
 
-    Ok(())
-}
+    fn push_checked(
+        entry: &mut String,
+        entries: &mut Vec<Value>,
+        elem_type: &Type,
+    ) -> Result<(), JsonConversionError> {
+        if !entry.is_empty() {
+            // While in usual postgres response we get nulls as None and everything else
+            // as Some(&str), in arrays we get NULL as unquoted 'NULL' string (while
+            // string with value 'NULL' will be represented by '"NULL"'). So catch NULLs
+            // here while we have quotation info and convert them to None.
+            if entry == "NULL" {
+                entries.push(pg_text_to_json(None, elem_type)?);
+            } else {
+                entries.push(pg_text_to_json(Some(entry), elem_type)?);
+            }
+            entry.clear();
+        }
 
-/// Parse postgres array into JSON array.
-///
-/// This is a bit involved because we need to handle nested arrays and quoted
-/// values. Unlike postgres we don't check that all nested arrays have the same
-/// dimensions, we just return them as is.
-///
-/// <https://www.postgresql.org/docs/current/arrays.html#ARRAYS-IO>
-///
-/// The external text representation of an array value consists of items that are interpreted
-/// according to the I/O conversion rules for the array's element type, plus decoration that
-/// indicates the array structure. The decoration consists of curly braces (`{` and `}`) around
-/// the array value plus delimiter characters between adjacent items. The delimiter character
-/// is usually a comma (,) but can be something else: it is determined by the typdelim setting
-/// for the array's element type. Among the standard data types provided in the PostgreSQL
-/// distribution, all use a comma, except for type box, which uses a semicolon (;).
-///
-/// In a multidimensional array, each dimension (row, plane, cube, etc.)
-/// gets its own level of curly braces, and delimiters must be written between adjacent
-/// curly-braced entities of the same level.
-fn pg_array_parse(
-    elements: &mut Vec<Value>,
-    mut pg_array: &str,
-    elem: &Type,
-    delim: char,
-) -> Result<(), JsonConversionError> {
-    // skip bounds decoration, eg:
-    // `[1:1][-2:-1][3:5]={{{1,2,3},{4,5,6}}}`
-    // technically these are significant, but we have no way to represent them in json.
-    if let Some('[') = pg_array.chars().next() {
-        let Some((_bounds, array)) = pg_array.split_once('=') else {
-            return Err(JsonConversionError::UnbalancedArray);
-        };
-        pg_array = array;
+        Ok(())
     }
 
-    // whitespace might preceed a `{`.
-    let pg_array = pg_array.trim_start();
+    while let Some((mut i, mut c)) = pg_array_chr.next() {
+        let mut escaped = false;
 
-    let rest = pg_array_parse_inner(elements, pg_array, elem, delim)?;
-    if !rest.is_empty() {
-        return Err(JsonConversionError::UnbalancedArray);
-    }
+        if c == '\\' {
+            escaped = true;
+            let Some(x) = pg_array_chr.next() else {
+                return Err(JsonConversionError::UnbalancedArray);
+            };
+            (i, c) = x;
+        }
 
-    Ok(())
-}
-
-/// reads a single array from the `pg_array` string and pushes each values to `elements`.
-/// returns the rest of the `pg_array` string that was not read.
-fn pg_array_parse_inner<'a>(
-    elements: &mut Vec<Value>,
-    mut pg_array: &'a str,
-    elem: &Type,
-    delim: char,
-) -> Result<&'a str, JsonConversionError> {
-    // array should have a `{` prefix.
-    pg_array = pg_array
-        .strip_prefix('{')
-        .ok_or(JsonConversionError::UnbalancedArray)?;
-
-    let mut q = String::new();
-
-    loop {
-        let value = push_entry(elements, Value::Null);
-        pg_array = pg_array_parse_item(value, &mut q, pg_array, elem, delim)?;
-
-        // check for separator.
-        if let Some(next) = pg_array.strip_prefix(delim) {
-            // next item.
-            pg_array = next;
-        } else {
-            break;
+        match c {
+            '{' if !quote => {
+                level += 1;
+                if level > 1 {
+                    let (res, off) = pg_array_parse_inner(&pg_array[i..], elem_type, true)?;
+                    entries.push(res);
+                    for _ in 0..off - 1 {
+                        pg_array_chr.next();
+                    }
+                }
+            }
+            '}' if !quote => {
+                level -= 1;
+                if level == 0 {
+                    push_checked(&mut entry, &mut entries, elem_type)?;
+                    if nested {
+                        return Ok((Value::Array(entries), i));
+                    }
+                }
+            }
+            '"' if !escaped => {
+                if quote {
+                    // end of quoted string, so push it manually without any checks
+                    // for emptiness or nulls
+                    entries.push(pg_text_to_json(Some(&entry), elem_type)?);
+                    entry.clear();
+                }
+                quote = !quote;
+            }
+            ',' if !quote => {
+                push_checked(&mut entry, &mut entries, elem_type)?;
+            }
+            _ => {
+                entry.push(c);
+            }
         }
     }
 
-    let Some(next) = pg_array.strip_prefix('}') else {
-        // missing `}` terminator.
+    if level != 0 {
         return Err(JsonConversionError::UnbalancedArray);
-    };
-
-    // whitespace might follow a `}`.
-    Ok(next.trim_start())
-}
-
-/// reads a single item from the `pg_array` string.
-/// returns the rest of the `pg_array` string that was not read.
-///
-/// `quoted` is a scratch allocation that has no defined output.
-fn pg_array_parse_item<'a>(
-    output: &mut Value,
-    quoted: &mut String,
-    mut pg_array: &'a str,
-    elem: &Type,
-    delim: char,
-) -> Result<&'a str, JsonConversionError> {
-    // We are trying to parse an array item.
-    // This could be a new array, if this is a multi-dimentional array.
-    // This could be a quoted string representing `elem`.
-    // This could be an unquoted string representing `elem`.
-
-    // whitespace might preceed an item.
-    pg_array = pg_array.trim_start();
-
-    if pg_array.starts_with('{') {
-        // nested array.
-        let mut nested = vec![];
-        pg_array = pg_array_parse_inner(&mut nested, pg_array, elem, delim)?;
-        *output = Value::Array(nested);
-        return Ok(pg_array);
     }
 
-    if let Some(mut pg_array) = pg_array.strip_prefix('"') {
-        // the parsed string is un-escaped and written into quoted.
-        pg_array = pg_array_parse_quoted(quoted, pg_array)?;
-
-        // we have un-escaped the string, parse it as pgtext.
-        pg_text_to_json(output, quoted, elem)?;
-
-        return Ok(pg_array);
-    }
-
-    // we need to parse an item. read until we find a delimiter or `}`.
-    let index = pg_array
-        .find([delim, '}'])
-        .ok_or(JsonConversionError::UnbalancedArray)?;
-
-    let item;
-    (item, pg_array) = pg_array.split_at(index);
-
-    // item might have trailing whitespace that we need to ignore.
-    let item = item.trim_end();
-
-    // we might have an item string:
-    // check for null
-    if item == "NULL" {
-        *output = Value::Null;
-    } else {
-        pg_text_to_json(output, item, elem)?;
-    }
-
-    Ok(pg_array)
-}
-
-/// reads a single quoted item from the `pg_array` string.
-///
-/// Returns the rest of the `pg_array` string that was not read.
-/// The output is written into `quoted`.
-///
-/// The pg_array string must have a `"` terminator, but the `"` initial value
-/// must have already been removed from the input. The terminator is removed.
-fn pg_array_parse_quoted<'a>(
-    quoted: &mut String,
-    mut pg_array: &'a str,
-) -> Result<&'a str, JsonConversionError> {
-    // The array output routine will put double quotes around element values if they are empty strings,
-    // contain curly braces, delimiter characters, double quotes, backslashes, or white space,
-    // or match the word `NULL`. Double quotes and backslashes embedded in element values will be backslash-escaped.
-    // For numeric data types it is safe to assume that double quotes will never appear,
-    // but for textual data types one should be prepared to cope with either the presence or absence of quotes.
-
-    quoted.clear();
-
-    // We write to quoted in chunks terminated by an escape character.
-    // Eg if we have the input `foo\"bar"`, then we write `foo`, then `"`, then finally `bar`.
-
-    loop {
-        // we need to parse an chunk. read until we find a '\\' or `"`.
-        let i = pg_array
-            .find(['\\', '"'])
-            .ok_or(JsonConversionError::UnbalancedString)?;
-
-        let chunk: &str;
-        (chunk, pg_array) = pg_array
-            .split_at_checked(i)
-            .expect("i is guaranteed to be in-bounds of pg_array");
-
-        // push the chunk.
-        quoted.push_str(chunk);
-
-        // consume the chunk_end character.
-        let chunk_end: char;
-        (chunk_end, pg_array) =
-            split_first_char(pg_array).expect("pg_array should start with either '\\\\' or '\"'");
-
-        // finished.
-        if chunk_end == '"' {
-            // whitespace might follow the '"'.
-            pg_array = pg_array.trim_start();
-
-            break Ok(pg_array);
-        }
-
-        // consume the escaped character.
-        let escaped: char;
-        (escaped, pg_array) =
-            split_first_char(pg_array).ok_or(JsonConversionError::UnbalancedString)?;
-
-        quoted.push(escaped);
-    }
-}
-
-fn split_first_char(s: &str) -> Option<(char, &str)> {
-    let mut chars = s.chars();
-    let c = chars.next()?;
-    Some((c, chars.as_str()))
+    Ok((Value::Array(entries), 0))
 }
 
 #[cfg(test)]
@@ -439,33 +316,37 @@ mod tests {
         );
     }
 
-    fn pg_text_to_json(val: &str, pg_type: &Type) -> Value {
-        let mut v = Value::Null;
-        super::pg_text_to_json(&mut v, val, pg_type).unwrap();
-        v
-    }
-
-    fn pg_array_parse(pg_array: &str, pg_type: &Type) -> Value {
-        let mut array = vec![];
-        super::pg_array_parse(&mut array, pg_array, pg_type, ',').unwrap();
-        Value::Array(array)
-    }
-
     #[test]
     fn test_atomic_types_parse() {
-        assert_eq!(pg_text_to_json("foo", &Type::TEXT), json!("foo"));
-        assert_eq!(pg_text_to_json("42", &Type::INT4), json!(42));
-        assert_eq!(pg_text_to_json("42", &Type::INT2), json!(42));
-        assert_eq!(pg_text_to_json("42", &Type::INT8), json!("42"));
-        assert_eq!(pg_text_to_json("42.42", &Type::FLOAT8), json!(42.42));
-        assert_eq!(pg_text_to_json("42.42", &Type::FLOAT4), json!(42.42));
-        assert_eq!(pg_text_to_json("NaN", &Type::FLOAT4), json!("NaN"));
         assert_eq!(
-            pg_text_to_json("Infinity", &Type::FLOAT4),
+            pg_text_to_json(Some("foo"), &Type::TEXT).unwrap(),
+            json!("foo")
+        );
+        assert_eq!(pg_text_to_json(None, &Type::TEXT).unwrap(), json!(null));
+        assert_eq!(pg_text_to_json(Some("42"), &Type::INT4).unwrap(), json!(42));
+        assert_eq!(pg_text_to_json(Some("42"), &Type::INT2).unwrap(), json!(42));
+        assert_eq!(
+            pg_text_to_json(Some("42"), &Type::INT8).unwrap(),
+            json!("42")
+        );
+        assert_eq!(
+            pg_text_to_json(Some("42.42"), &Type::FLOAT8).unwrap(),
+            json!(42.42)
+        );
+        assert_eq!(
+            pg_text_to_json(Some("42.42"), &Type::FLOAT4).unwrap(),
+            json!(42.42)
+        );
+        assert_eq!(
+            pg_text_to_json(Some("NaN"), &Type::FLOAT4).unwrap(),
+            json!("NaN")
+        );
+        assert_eq!(
+            pg_text_to_json(Some("Infinity"), &Type::FLOAT4).unwrap(),
             json!("Infinity")
         );
         assert_eq!(
-            pg_text_to_json("-Infinity", &Type::FLOAT4),
+            pg_text_to_json(Some("-Infinity"), &Type::FLOAT4).unwrap(),
             json!("-Infinity")
         );
 
@@ -474,9 +355,10 @@ mod tests {
                 .unwrap();
         assert_eq!(
             pg_text_to_json(
-                r#"{"s":"str","n":42,"f":4.2,"a":[null,3,"a"]}"#,
+                Some(r#"{"s":"str","n":42,"f":4.2,"a":[null,3,"a"]}"#),
                 &Type::JSONB
-            ),
+            )
+            .unwrap(),
             json
         );
     }
@@ -484,7 +366,7 @@ mod tests {
     #[test]
     fn test_pg_array_parse_text() {
         fn pt(pg_arr: &str) -> Value {
-            pg_array_parse(pg_arr, &Type::TEXT)
+            pg_array_parse(pg_arr, &Type::TEXT).unwrap()
         }
         assert_eq!(
             pt(r#"{"aa\"\\\,a",cha,"bbbb"}"#),
@@ -507,7 +389,7 @@ mod tests {
     #[test]
     fn test_pg_array_parse_bool() {
         fn pb(pg_arr: &str) -> Value {
-            pg_array_parse(pg_arr, &Type::BOOL)
+            pg_array_parse(pg_arr, &Type::BOOL).unwrap()
         }
         assert_eq!(pb(r#"{t,f,t}"#), json!([true, false, true]));
         assert_eq!(pb(r#"{{t,f,t}}"#), json!([[true, false, true]]));
@@ -524,7 +406,7 @@ mod tests {
     #[test]
     fn test_pg_array_parse_numbers() {
         fn pn(pg_arr: &str, ty: &Type) -> Value {
-            pg_array_parse(pg_arr, ty)
+            pg_array_parse(pg_arr, ty).unwrap()
         }
         assert_eq!(pn(r#"{1,2,3}"#, &Type::INT4), json!([1, 2, 3]));
         assert_eq!(pn(r#"{1,2,3}"#, &Type::INT2), json!([1, 2, 3]));
@@ -552,7 +434,7 @@ mod tests {
     #[test]
     fn test_pg_array_with_decoration() {
         fn p(pg_arr: &str) -> Value {
-            pg_array_parse(pg_arr, &Type::INT2)
+            pg_array_parse(pg_arr, &Type::INT2).unwrap()
         }
         assert_eq!(
             p(r#"[1:1][-2:-1][3:5]={{{1,2,3},{4,5,6}}}"#),
@@ -563,7 +445,7 @@ mod tests {
     #[test]
     fn test_pg_array_parse_json() {
         fn pt(pg_arr: &str) -> Value {
-            pg_array_parse(pg_arr, &Type::JSONB)
+            pg_array_parse(pg_arr, &Type::JSONB).unwrap()
         }
         assert_eq!(pt(r#"{"{}"}"#), json!([{}]));
         assert_eq!(

proxy/src/serverless/sql_over_http.rs

@@ -22,7 +22,7 @@ use serde_json::Value;
 use serde_json::value::RawValue;
 use tokio::time::{self, Instant};
 use tokio_util::sync::CancellationToken;
-use tracing::{Level, debug, error, info};
+use tracing::{debug, error, info};
 use typed_json::json;
 use url::Url;
 use uuid::Uuid;
@@ -33,9 +33,9 @@ use super::conn_pool_lib::{self, ConnInfo};
 use super::error::HttpCodeError;
 use super::http_util::json_response;
 use super::json::{JsonConversionError, json_to_pg_text, pg_text_row_to_json};
-use crate::auth::ComputeUserInfoParseError;
 use crate::auth::backend::{ComputeCredentialKeys, ComputeUserInfo};
-use crate::config::{AuthenticationConfig, HttpConfig, ProxyConfig};
+use crate::auth::{ComputeUserInfoParseError, endpoint_sni};
+use crate::config::{AuthenticationConfig, HttpConfig, ProxyConfig, TlsConfig};
 use crate::context::RequestContext;
 use crate::error::{ErrorKind, ReportableError, UserFacingError};
 use crate::http::{ReadBodyError, read_body_with_limit};
@@ -43,7 +43,7 @@ use crate::metrics::{HttpDirection, Metrics, SniGroup, SniKind};
 use crate::pqproto::StartupMessageParams;
 use crate::proxy::NeonOptions;
 use crate::serverless::backend::HttpConnError;
-use crate::types::{DbName, EndpointId, RoleName};
+use crate::types::{DbName, RoleName};
 use crate::usage_metrics::{MetricCounter, MetricCounterRecorder};
 use crate::util::run_until_cancelled;
 
@@ -113,6 +113,8 @@ pub(crate) enum ConnInfoError {
     MissingHostname,
     #[error("invalid hostname: {0}")]
     InvalidEndpoint(#[from] ComputeUserInfoParseError),
+    #[error("malformed endpoint")]
+    MalformedEndpoint,
 }
 
 #[derive(Debug, thiserror::Error)]
@@ -139,6 +141,7 @@ fn get_conn_info(
     config: &'static AuthenticationConfig,
     ctx: &RequestContext,
     headers: &HeaderMap,
+    tls: Option<&TlsConfig>,
 ) -> Result<ConnInfoWithAuth, ConnInfoError> {
     let connection_string = headers
         .get(&CONN_STRING)
@@ -196,11 +199,17 @@ fn get_conn_info(
         return Err(ConnInfoError::MissingCredentials(Credentials::Password));
     };
 
-    let endpoint: EndpointId = match connection_url.host() {
-        Some(url::Host::Domain(hostname)) => hostname
-            .split_once('.')
-            .map_or(hostname, |(prefix, _)| prefix)
-            .into(),
+    let endpoint = match connection_url.host() {
+        Some(url::Host::Domain(hostname)) => {
+            if let Some(tls) = tls {
+                endpoint_sni(hostname, &tls.common_names).ok_or(ConnInfoError::MalformedEndpoint)?
+            } else {
+                hostname
+                    .split_once('.')
+                    .map_or(hostname, |(prefix, _)| prefix)
+                    .into()
+            }
+        }
         Some(url::Host::Ipv4(_) | url::Host::Ipv6(_)) | None => {
             return Err(ConnInfoError::MissingHostname);
         }
@@ -381,35 +390,12 @@ pub(crate) async fn handle(
             let line = get(db_error, |db| db.line().map(|l| l.to_string()));
             let routine = get(db_error, |db| db.routine());
 
-            match &e {
-                SqlOverHttpError::Postgres(e)
-                    if e.as_db_error().is_some() && error_kind == ErrorKind::User =>
-                {
-                    // this error contains too much info, and it's not an error we care about.
-                    if tracing::enabled!(Level::DEBUG) {
-                        tracing::debug!(
-                            kind=error_kind.to_metric_label(),
-                            error=%e,
-                            msg=message,
-                            "forwarding error to user"
-                        );
-                    } else {
-                        tracing::info!(
-                            kind = error_kind.to_metric_label(),
-                            error = "bad query",
-                            "forwarding error to user"
-                        );
-                    }
-                }
-                _ => {
-                    tracing::info!(
-                        kind=error_kind.to_metric_label(),
-                        error=%e,
-                        msg=message,
-                        "forwarding error to user"
-                    );
-                }
-            }
+            tracing::info!(
+                kind=error_kind.to_metric_label(),
+                error=%e,
+                msg=message,
+                "forwarding error to user"
+            );
 
             json_response(
                 e.get_http_status_code(),
@@ -474,15 +460,7 @@ impl ReportableError for SqlOverHttpError {
             SqlOverHttpError::ConnInfo(e) => e.get_error_kind(),
             SqlOverHttpError::ResponseTooLarge(_) => ErrorKind::User,
             SqlOverHttpError::InvalidIsolationLevel => ErrorKind::User,
-            // customer initiated SQL errors.
-            SqlOverHttpError::Postgres(p) => {
-                if p.as_db_error().is_some() {
-                    ErrorKind::User
-                } else {
-                    ErrorKind::Compute
-                }
-            }
-            // proxy initiated SQL errors.
+            SqlOverHttpError::Postgres(p) => p.get_error_kind(),
             SqlOverHttpError::InternalPostgres(p) => {
                 if p.as_db_error().is_some() {
                     ErrorKind::Service
@@ -490,7 +468,6 @@ impl ReportableError for SqlOverHttpError {
                     ErrorKind::Compute
                 }
             }
-            // postgres returned a bad row format that we couldn't parse.
             SqlOverHttpError::JsonConversion(_) => ErrorKind::Postgres,
             SqlOverHttpError::Cancelled(c) => c.get_error_kind(),
         }
@@ -661,7 +638,14 @@ async fn handle_inner(
         "handling interactive connection from client"
     );
 
-    let conn_info = get_conn_info(&config.authentication_config, ctx, request.headers())?;
+    let conn_info = get_conn_info(
+        &config.authentication_config,
+        ctx,
+        request.headers(),
+        // todo: race condition?
+        // we're unlikely to change the common names.
+        config.tls_config.load().as_deref(),
+    )?;
     info!(
         user = conn_info.conn_info.user_info.user.as_str(),
         "credentials"
@@ -1119,6 +1103,7 @@ async fn query_to_json<T: GenericClient>(
 
     let columns_len = row_stream.statement.columns().len();
     let mut fields = Vec::with_capacity(columns_len);
+    let mut types = Vec::with_capacity(columns_len);
 
     for c in row_stream.statement.columns() {
         fields.push(json!({
@@ -1130,6 +1115,8 @@ async fn query_to_json<T: GenericClient>(
             "dataTypeModifier": c.type_modifier(),
             "format": "text",
         }));
+
+        types.push(c.type_().clone());
     }
 
     let raw_output = parsed_headers.raw_output;
@@ -1151,7 +1138,7 @@ async fn query_to_json<T: GenericClient>(
             ));
         }
 
-        let row = pg_text_row_to_json(&row, raw_output, array_mode)?;
+        let row = pg_text_row_to_json(&row, &types, raw_output, array_mode)?;
         rows.push(row);
 
         // assumption: parsing pg text and converting to json takes CPU time.

safekeeper/client/src/mgmt_api.rs

@@ -52,7 +52,7 @@ pub trait ResponseErrorMessageExt: Sized {
 impl ResponseErrorMessageExt for reqwest::Response {
     async fn error_from_body(self) -> Result<Self> {
         let status = self.status();
-        if status.is_success() {
+        if !(status.is_client_error() || status.is_server_error()) {
             return Ok(self);
         }
 

test_runner/fixtures/neon_fixtures.py

@@ -724,21 +724,15 @@ class NeonEnvBuilder:
 
         shutil.copytree(storcon_db_from_dir, storcon_db_to_dir, ignore=ignore_postgres_log)
         assert not (storcon_db_to_dir / "postgres.log").exists()
-
         # NB: neon_local rewrites postgresql.conf on each start based on neon_local config. No need to patch it.
-        # However, in this new NeonEnv, the pageservers and safekeepers listen on different ports, and the storage
-        # controller will currently reject re-attach requests from them because the NodeMetadata isn't identical.
+        # However, in this new NeonEnv, the pageservers listen on different ports, and the storage controller
+        # will currently reject re-attach requests from them because the NodeMetadata isn't identical.
         # So, from_repo_dir patches up the the storcon database.
         patch_script_path = self.repo_dir / "storage_controller_db.startup.sql"
         assert not patch_script_path.exists()
         patch_script = ""
-
         for ps in self.env.pageservers:
-            patch_script += f"UPDATE nodes SET listen_http_port={ps.service_port.http}, listen_pg_port={ps.service_port.pg}  WHERE node_id = '{ps.id}';\n"
-
-        for sk in self.env.safekeepers:
-            patch_script += f"UPDATE safekeepers SET http_port={sk.port.http}, port={sk.port.pg} WHERE id = '{sk.id}';\n"
-
+            patch_script += f"UPDATE nodes SET listen_http_port={ps.service_port.http}, listen_pg_port={ps.service_port.pg}  WHERE node_id = '{ps.id}';"
         patch_script_path.write_text(patch_script)
 
         # Update the config with info about tenants and timelines

test_runner/regress/test_compatibility.py

@@ -76,7 +76,6 @@ if TYPE_CHECKING:
 #    export CHECK_ONDISK_DATA_COMPATIBILITY=true
 #    export COMPATIBILITY_NEON_BIN=neon_previous/target/${BUILD_TYPE}
 #    export COMPATIBILITY_POSTGRES_DISTRIB_DIR=neon_previous/pg_install
-#    export COMPATIBILITY_SNAPSHOT_DIR=test_output/compatibility_snapshot_pgv${DEFAULT_PG_VERSION}
 #
 #    # Build previous version of binaries and store them somewhere:
 #    rm -rf pg_install target
@@ -103,7 +102,6 @@ if TYPE_CHECKING:
 #    export CHECK_ONDISK_DATA_COMPATIBILITY=true
 #    export COMPATIBILITY_NEON_BIN=neon_previous/target/${BUILD_TYPE}
 #    export COMPATIBILITY_POSTGRES_DISTRIB_DIR=neon_previous/pg_install
-#    export COMPATIBILITY_SNAPSHOT_DIR=test_output/compatibility_snapshot_pgv${DEFAULT_PG_VERSION}
 #    export NEON_BIN=target/${BUILD_TYPE}
 #    export POSTGRES_DISTRIB_DIR=pg_install
 #

test_runner/regress/test_pageserver_metric_collection.py

@@ -180,7 +180,7 @@ def test_metric_collection(
     httpserver.check()
 
     # Check that at least one bucket output object is present, and that all
-    # can be decompressed and decoded as NDJSON.
+    # can be decompressed and decoded.
     bucket_dumps = {}
     assert isinstance(env.pageserver_remote_storage, LocalFsStorage)
     for dirpath, _dirs, files in os.walk(env.pageserver_remote_storage.root):
@@ -188,13 +188,7 @@ def test_metric_collection(
             file_path = os.path.join(dirpath, file)
             log.info(file_path)
             if file.endswith(".gz"):
-                events = []
-                with gzip.open(file_path, "rt") as f:
-                    for line in f:
-                        line = line.strip()
-                        if line:
-                            events.append(json.loads(line))
-                bucket_dumps[file_path] = {"events": events}
+                bucket_dumps[file_path] = json.load(gzip.open(file_path))
 
     assert len(bucket_dumps) >= 1
     assert all("events" in data for data in bucket_dumps.values())

test_runner/regress/test_timeline_detach_ancestor.py

@@ -324,7 +324,7 @@ def test_ancestor_detach_reparents_earlier(neon_env_builder: NeonEnvBuilder):
             # it is to be in line with the deletion timestamp.. well, almost.
             when = original_ancestor[2][:26]
             when_ts = datetime.datetime.fromisoformat(when).replace(tzinfo=datetime.UTC)
-            now = datetime.datetime.now(datetime.UTC)
+            now = datetime.datetime.utcnow().replace(tzinfo=datetime.UTC)
             assert when_ts < now
             assert len(lineage.get("reparenting_history", [])) == 0
         elif expected_ancestor == timeline_id:
@@ -458,20 +458,19 @@ def test_ancestor_detach_behavior_v2(neon_env_builder: NeonEnvBuilder, snapshots
 
     env.pageserver.quiesce_tenants()
 
-    # checking the ancestor after is much faster than waiting for the endpoint to start
+    # checking the ancestor after is much faster than waiting for the endpoint not start
     expected_result = [
-        # (branch_name, queried_timeline, expected_ancestor, rows, starts, read_only)
-        ("main", env.initial_timeline, None, 24576, 1, False),
-        ("after", after, env.initial_timeline, 24576, 1, False),
-        ("snapshot_branchpoint_old", snapshot_branchpoint_old, env.initial_timeline, 8192, 1, True),
-        ("snapshot_branchpoint", snapshot_branchpoint, env.initial_timeline, 16384, 1, False),
-        ("branch_to_detach", branch_to_detach, None, 16384, 1, False),
-        ("earlier", earlier, env.initial_timeline, 0, 1, False),
+        ("main", env.initial_timeline, None, 24576, 1),
+        ("after", after, env.initial_timeline, 24576, 1),
+        ("snapshot_branchpoint_old", snapshot_branchpoint_old, env.initial_timeline, 8192, 1),
+        ("snapshot_branchpoint", snapshot_branchpoint, env.initial_timeline, 16384, 1),
+        ("branch_to_detach", branch_to_detach, None, 16384, 1),
+        ("earlier", earlier, env.initial_timeline, 0, 1),
     ]
 
     assert isinstance(env.pageserver_remote_storage, LocalFsStorage)
 
-    for branch_name, queried_timeline, expected_ancestor, _, _, _ in expected_result:
+    for branch_name, queried_timeline, expected_ancestor, _, _ in expected_result:
         details = client.timeline_detail(env.initial_tenant, queried_timeline)
         ancestor_timeline_id = details["ancestor_timeline_id"]
         if expected_ancestor is None:
@@ -509,17 +508,13 @@ def test_ancestor_detach_behavior_v2(neon_env_builder: NeonEnvBuilder, snapshots
             assert len(lineage.get("original_ancestor", [])) == 0
             assert len(lineage.get("reparenting_history", [])) == 0
 
-    for branch_name, queried_timeline, _, rows, starts, read_only in expected_result:
-        last_record_lsn = None
-        if read_only:
-            # specifying the lsn makes the endpoint read-only and not connect to safekeepers
-            details = client.timeline_detail(env.initial_tenant, queried_timeline)
-            last_record_lsn = Lsn(details["last_record_lsn"])
-
-        log.info(f"reading data from branch {branch_name} at {last_record_lsn}")
+    for branch_name, queried_timeline, _, rows, starts in expected_result:
+        details = client.timeline_detail(env.initial_tenant, queried_timeline)
+        log.info(f"reading data from branch {branch_name}")
+        # specifying the lsn makes the endpoint read-only and not connect to safekeepers
         with env.endpoints.create(
             branch_name,
-            lsn=last_record_lsn,
+            lsn=Lsn(details["last_record_lsn"]),
         ) as ep:
             ep.start(safekeeper_generation=1)
             assert ep.safe_psql("SELECT count(*) FROM foo;")[0][0] == rows