diff --git a/tensorstore/kvstore/neuroglancer_uint64_sharded/neuroglancer_uint64_sharded.cc b/tensorstore/kvstore/neuroglancer_uint64_sharded/neuroglancer_uint64_sharded.cc
index dd80c0623..22d5298f0 100644
--- a/tensorstore/kvstore/neuroglancer_uint64_sharded/neuroglancer_uint64_sharded.cc
+++ b/tensorstore/kvstore/neuroglancer_uint64_sharded/neuroglancer_uint64_sharded.cc
@@ -584,8 +584,8 @@ class ShardedKeyValueStoreWriteCache
           entry, std::move(read_result));
     }
 
-    absl::Time apply_staleness_bound_;
     ApplyReceiver apply_receiver_;
+    ApplyOptions apply_options_;
     absl::Status apply_status_;
   };
 
@@ -645,7 +645,7 @@ void ShardedKeyValueStoreWriteCache::TransactionNode::WritebackError() {
 
 namespace {
 void StartApply(ShardedKeyValueStoreWriteCache::TransactionNode& node) {
-  RetryAtomicWriteback(node.phases_, node.apply_staleness_bound_);
+  RetryAtomicWriteback(node.phases_, node.apply_options_.staleness_bound);
 }
 
 /// Attempts to compute the new encoded shard state that merges any mutations
@@ -747,7 +747,7 @@ void MergeForWriteback(ShardedKeyValueStoreWriteCache::TransactionNode& node,
     // However, even in that case, the extra retries aren't really an added
     // cost, because those retries would still be needed anyway due to the
     // concurrent modifications.
-    node.apply_staleness_bound_ = absl::Now();
+    node.apply_options_.staleness_bound = absl::Now();
     StartApply(node);
     return;
   }
@@ -769,7 +769,7 @@ void MergeForWriteback(ShardedKeyValueStoreWriteCache::TransactionNode& node,
 void ShardedKeyValueStoreWriteCache::TransactionNode::DoApply(
     ApplyOptions options, ApplyReceiver receiver) {
   apply_receiver_ = std::move(receiver);
-  apply_staleness_bound_ = options.staleness_bound;
+  apply_options_ = options;
   apply_status_ = absl::Status();
 
   GetOwningCache(*this).executor()([this] { StartApply(*this); });
@@ -786,6 +786,7 @@ void ShardedKeyValueStoreWriteCache::TransactionNode::AllEntriesDone(
   GetOwningCache(*this).executor()([&self] {
     TimestampedStorageGeneration stamp;
     bool mismatch = false;
+    bool modified = false;
 
     size_t num_chunks = 0;
 
@@ -795,8 +796,12 @@ void ShardedKeyValueStoreWriteCache::TransactionNode::AllEntriesDone(
       auto& buffered_entry =
           static_cast<AtomicMultiPhaseMutation::BufferedReadModifyWriteEntry&>(
               entry);
+      if (buffered_entry.read_result_.state !=
+          kvstore::ReadResult::kUnspecified) {
+        modified = true;
+        ++num_chunks;
+      }
       auto& entry_stamp = buffered_entry.read_result_.stamp;
-      ++num_chunks;
       if (StorageGeneration::IsConditional(entry_stamp.generation)) {
         if (!StorageGeneration::IsUnknown(stamp.generation) &&
             StorageGeneration::Clean(stamp.generation) !=
@@ -810,17 +815,26 @@ void ShardedKeyValueStoreWriteCache::TransactionNode::AllEntriesDone(
     }
 
     if (mismatch) {
-      self.apply_staleness_bound_ = absl::Now();
+      self.apply_options_.staleness_bound = absl::Now();
       StartApply(self);
       return;
     }
     auto& cache = GetOwningCache(self);
+    if (!modified && StorageGeneration::IsUnknown(stamp.generation) &&
+        self.apply_options_.apply_mode !=
+            ApplyOptions::ApplyMode::kSpecifyUnchanged) {
+      internal::AsyncCache::ReadState update;
+      update.stamp = TimestampedStorageGeneration::Unconditional();
+      execution::set_value(std::exchange(self.apply_receiver_, {}),
+                           std::move(update));
+      return;
+    }
     if (!StorageGeneration::IsUnknown(stamp.generation) ||
         !cache.get_max_chunks_per_shard_ ||
         cache.get_max_chunks_per_shard_(GetOwningEntry(self).shard()) !=
             num_chunks) {
       self.internal::AsyncCache::TransactionNode::Read(
-              self.apply_staleness_bound_)
+              self.apply_options_.staleness_bound)
           .ExecuteWhenReady([&self](ReadyFuture<const void> future) {
             if (!future.result().ok()) {
               execution::set_error(std::exchange(self.apply_receiver_, {}),
diff --git a/tensorstore/kvstore/neuroglancer_uint64_sharded/neuroglancer_uint64_sharded_test.cc b/tensorstore/kvstore/neuroglancer_uint64_sharded/neuroglancer_uint64_sharded_test.cc
index 33a515ca7..9bc7a9644 100644
--- a/tensorstore/kvstore/neuroglancer_uint64_sharded/neuroglancer_uint64_sharded_test.cc
+++ b/tensorstore/kvstore/neuroglancer_uint64_sharded/neuroglancer_uint64_sharded_test.cc
@@ -892,6 +892,25 @@ TEST_F(UnderlyingKeyValueStoreTest, Read) {
   }
 }
 
+// Verify that a read-only transaction does not do any I/O on commit.
+TEST_F(UnderlyingKeyValueStoreTest, TransactionReadThenCommit) {
+  tensorstore::Transaction txn(tensorstore::isolated);
+  auto memory_store = tensorstore::GetMemoryKeyValueStore();
+  {
+    auto future = kvstore::Read(KvStore{store, txn}, GetChunkKey(0x50), {});
+    {
+      auto req = mock_store->read_requests.pop();
+      req(memory_store);
+      ASSERT_EQ(0, mock_store->read_requests.size());
+    }
+    EXPECT_THAT(future.result(),
+                ::testing::Optional(MatchesKvsReadResultNotFound()));
+  }
+
+  auto commit_future = txn.CommitAsync();
+  TENSORSTORE_ASSERT_OK(commit_future.result());
+}
+
 // Tests issuing read for chunk in uncached minishard index while there is a
 // concurrent modification.
 TEST_F(UnderlyingKeyValueStoreTest,