Defer eviction of checkpoint blocks to reduce post-checkpoint cold reads

src/include/duckdb/storage/block_manager.hpp

@@ -13,6 +13,7 @@
 #include "duckdb/common/optional_idx.hpp"
 #include "duckdb/common/unordered_map.hpp"
 #include "duckdb/storage/block.hpp"
+#include "duckdb/storage/buffer/buffer_handle.hpp"
 #include "duckdb/storage/storage_info.hpp"
 
 namespace duckdb {
@@ -182,6 +183,12 @@ class BlockManager {
 		return reinterpret_cast<const TARGET &>(*this);
 	}
 
+public:
+	//! Begin deferring eviction queue additions (call before checkpoint writes)
+	void BeginDeferEviction();
+	//! End deferring and release all retained blocks to the eviction queue
+	void EndDeferEviction();
+
 protected:
 	//! A flag to be flipped in the destructor of the subclass, which is called first.
 	//! Relevant for some Windows edge cases.
@@ -192,6 +199,10 @@ class BlockManager {
 	mutex blocks_lock;
 	//! A mapping of block id -> BlockHandle
 	unordered_map<block_id_t, weak_ptr<BlockHandle>> blocks;
+	//! When true, ConvertToPersistent pins blocks instead of adding to eviction queue
+	bool defer_eviction = false;
+	//! Blocks pinned during deferred eviction (kept alive until EndDeferEviction)
+	vector<BufferHandle> deferred_pins;
 	//! The metadata manager
 	unique_ptr<MetadataManager> metadata_manager;
 	//! The allocation size of blocks managed by this block manager. Defaults to DEFAULT_BLOCK_ALLOC_SIZE

src/storage/buffer/block_manager.cpp

@@ -104,10 +104,13 @@ shared_ptr<BlockHandle> BlockManager::ConvertToPersistent(QueryContext context,
 	old_handle.Destroy();
 	old_block.reset();
 
-	// potentially purge the queue
-	auto purge_queue = buffer_manager.GetBufferPool().AddToEvictionQueue(new_block);
-	if (purge_queue) {
-		buffer_manager.GetBufferPool().PurgeQueue(*new_block);
+	if (defer_eviction) {
+		deferred_pins.push_back(buffer_manager.Pin(new_block));
+	} else {
+		auto purge_queue = buffer_manager.GetBufferPool().AddToEvictionQueue(new_block);
+		if (purge_queue) {
+			buffer_manager.GetBufferPool().PurgeQueue(*new_block);
+		}
 	}
 	return new_block;
 }
@@ -148,4 +151,17 @@ void BlockManager::Write(QueryContext context, FileBuffer &block, block_id_t blo
 void BlockManager::Truncate() {
 }
 
+void BlockManager::BeginDeferEviction() {
+	defer_eviction = true;
+}
+
+void BlockManager::EndDeferEviction() {
+	defer_eviction = false;
+	// Release all pins — blocks become eligible for eviction
+	for (auto &pin : deferred_pins) {
+		pin.Destroy();
+	}
+	deferred_pins.clear();
+}
+
 } // namespace duckdb

src/storage/checkpoint_manager.cpp

@@ -247,6 +247,7 @@ void SingleFileCheckpointWriter::CreateCheckpoint() {
 		dependency_manager.ReorderEntries(catalog_entries);
 
 		// write the actual data into the database
+		block_manager.BeginDeferEviction();
 
 		// Create a serializer to write the checkpoint data
 		// The serialized format is roughly:
@@ -277,6 +278,8 @@ void SingleFileCheckpointWriter::CreateCheckpoint() {
 		metadata_writer->Flush();
 		table_metadata_writer->Flush();
 
+		block_manager.EndDeferEviction();
+
 		auto debug_checkpoint_abort = Settings::Get<DebugCheckpointAbortSetting>(db.GetDatabase());
 		if (debug_checkpoint_abort == CheckpointAbort::DEBUG_ABORT_BEFORE_HEADER) {
 			throw FatalException("Checkpoint aborted before header write because of PRAGMA checkpoint_abort flag");