Properly set upper bound of subcompaction output (#4879) (#4898)

Summary:
Fix the ouput overlap bug when using subcompactions, the upper bound of output
file was extended incorrectly.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4898

Differential Revision: D13736107

Pulled By: ajkr

fbshipit-source-id: 21dca09f81d5f07bf2766bf566f9b50dcab7d8e3

diff --git a/db/compaction_job.cc b/db/compaction_job.cc
index 0bdf78cfce443ae43e5e49751adece0608316d83..10aaef0983dc488e7cfb4bc6f85ab63bf16b9b3e 100644 (file)
--- a/db/compaction_job.cc
+++ b/db/compaction_job.cc
@@ -1204,10 +1204,19 @@ Status CompactionJob::FinishCompactionOutputFile(
       lower_bound = nullptr;
     }
     if (next_table_min_key != nullptr) {
-      // This isn't the last file in the subcompaction, so extend until the next
-      // file starts.
+      // This may be the last file in the subcompaction in some cases, so we
+      // need to compare the end key of subcompaction with the next file start
+      // key. When the end key is chosen by the subcompaction, we know that
+      // it must be the biggest key in output file. Therefore, it is safe to
+      // use the smaller key as the upper bound of the output file, to ensure
+      // that there is no overlapping between different output files.
       upper_bound_guard = ExtractUserKey(*next_table_min_key);
-      upper_bound = &upper_bound_guard;
+      if (sub_compact->end != nullptr &&
+          ucmp->Compare(upper_bound_guard, *sub_compact->end) >= 0) {
+        upper_bound = sub_compact->end;
+      } else {
+        upper_bound = &upper_bound_guard;
+      }
     } else {
       // This is the last file in the subcompaction, so extend until the
       // subcompaction ends.
@@ -1225,6 +1234,13 @@ Status CompactionJob::FinishCompactionOutputFile(
       has_overlapping_endpoints = false;
     }
 
+    // The end key of the subcompaction must be bigger or equal to the upper
+    // bound. If the end of subcompaction is null or the upper bound is null,
+    // it means that this file is the last file in the compaction. So there
+    // will be no overlapping between this file and others.
+    assert(sub_compact->end == nullptr ||
+           upper_bound == nullptr ||
+           ucmp->Compare(*upper_bound , *sub_compact->end) <= 0);
     auto it = range_del_agg->NewIterator(lower_bound, upper_bound,
                                          has_overlapping_endpoints);
     // Position the range tombstone output iterator. There may be tombstone