Fix handling of write intents for aborted transaction

Summary:
Transaction could be aborted while we are trying to write intents for it.
So we should handle the case when metadata is not present for transaction.
Also register request to be able to update last write id when operation finishes.

Test Plan: ybd --java-test org.yb.cql.TestReturnsClause#testBatchDMLWithUniqueIndex

Reviewers: mikhail, venkatesh, timur, robert

Reviewed By: robert

Subscribers: ybase

Differential Revision: https://phabricator.dev.yugabyte.com/D5302

src/yb/tablet/tablet.cc

@@ -640,14 +640,21 @@ void Tablet::PrepareTransactionWriteBatch(
     const KeyValueWriteBatchPB& put_batch,
     HybridTime hybrid_time,
     rocksdb::WriteBatch* rocksdb_write_batch) {
+  RequestScope request_scope(transaction_participant_.get());
   if (put_batch.transaction().has_isolation()) {
     // Store transaction metadata (status tablet, isolation level etc.)
     transaction_participant()->Add(put_batch.transaction(), rocksdb_write_batch);
   }
   auto transaction_id = CHECK_RESULT(
       FullyDecodeTransactionId(put_batch.transaction().transaction_id()));
   auto metadata_with_write_id = transaction_participant()->MetadataWithWriteId(transaction_id);
-  CHECK(metadata_with_write_id) << "Transaction metadata missing: " << transaction_id;
+  if (!metadata_with_write_id) {
+    // If metadata is missing it could be caused by aborted and removed transaction.
+    // In this case we should not add new intents for it.
+    LOG(INFO) << "Transaction metadata missing: " << transaction_id
+              << ", looks like it was just aborted";
+    return;
+  }
 
   auto isolation_level = metadata_with_write_id->first.isolation;
   auto write_id = metadata_with_write_id->second;