From 50994bf6990eae4155f7541161f5eaf1edb3ef08 Mon Sep 17 00:00:00 2001
From: Mark Callaghan <mdcallag@gmail.com>
Date: Thu, 19 Dec 2013 10:02:53 -0800
Subject: [PATCH] Don't always compress L0 files written by memtable flush

Summary:
Code was always compressing L0 files written by a memtable flush
when compression was enabled. Now this is done when
min_level_to_compress=0 for leveled compaction and when
universal_compaction_size_percent=-1 for universal compaction.

Task ID: #3416472

Blame Rev:

Test Plan:
ran db_bench with compression options

Revert Plan:

Database Impact:

Memcache Impact:

Other Notes:

EImportant:

- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -

Reviewers: dhruba, igor, sdong

Reviewed By: dhruba

CC: leveldb

Differential Revision: https://reviews.facebook.net/D14757
---
 db/builder.cc |  4 ++--
 db/builder.h  |  2 +-
 db/db_impl.cc | 33 ++++++++++++++++++++++++++-------
 db/db_impl.h  |  3 +++
 db/repair.cc  |  3 ++-
 5 files changed, 34 insertions(+), 11 deletions(-)

diff --git a/db/builder.cc b/db/builder.cc
index ad1334a15..61671db0d 100644
--- a/db/builder.cc
+++ b/db/builder.cc
@@ -42,7 +42,7 @@ Status BuildTable(const std::string& dbname,
                   const Comparator* user_comparator,
                   const SequenceNumber newest_snapshot,
                   const SequenceNumber earliest_seqno_in_memtable,
-                  const bool enable_compression) {
+                  const CompressionType compression) {
   Status s;
   meta->file_size = 0;
   meta->smallest_seqno = meta->largest_seqno = 0;
@@ -65,7 +65,7 @@ Status BuildTable(const std::string& dbname,
     }
 
     TableBuilder* builder = GetTableBuilder(options, file.get(),
-                                            options.compression);
+                                            compression);
 
     // the first key is the smallest key
     Slice key = iter->key();
diff --git a/db/builder.h b/db/builder.h
index 8c525bd05..2600dc24b 100644
--- a/db/builder.h
+++ b/db/builder.h
@@ -43,6 +43,6 @@ extern Status BuildTable(const std::string& dbname,
                          const Comparator* user_comparator,
                          const SequenceNumber newest_snapshot,
                          const SequenceNumber earliest_seqno_in_memtable,
-                         const bool enable_compression);
+                         const CompressionType compression);
 
 }  // namespace rocksdb
diff --git a/db/db_impl.cc b/db/db_impl.cc
index b8ae72fd9..169556e23 100644
--- a/db/db_impl.cc
+++ b/db/db_impl.cc
@@ -225,6 +225,28 @@ CompressionType GetCompressionType(const Options& options, int level,
   }
 }
 
+CompressionType GetCompressionFlush(const Options& options) {
+  // Compressing memtable flushes might not help unless the sequential load
+  // optimization is used for leveled compaction. Otherwise the CPU and
+  // latency overhead is not offset by saving much space.
+
+  bool can_compress;
+
+  if  (options.compaction_style == kCompactionStyleUniversal) {
+    can_compress =
+        (options.compaction_options_universal.compression_size_percent < 0);
+  } else {
+    // For leveled compress when min_level_to_compress == 0.
+    can_compress = (GetCompressionType(options, 0, true) != kNoCompression);
+  }
+
+  if (can_compress) {
+    return options.compression;
+  } else {
+    return kNoCompression;
+  }
+}
+
 DBImpl::DBImpl(const Options& options, const std::string& dbname)
     : env_(options.env),
       dbname_(dbname),
@@ -1068,7 +1090,8 @@ Status DBImpl::WriteLevel0TableForRecovery(MemTable* mem, VersionEdit* edit) {
     s = BuildTable(dbname_, env_, options_, storage_options_,
                    table_cache_.get(), iter, &meta,
                    user_comparator(), newest_snapshot,
-                   earliest_seqno_in_memtable, true);
+                   earliest_seqno_in_memtable,
+                   GetCompressionFlush(options_));
     LogFlush(options_.info_log);
     mutex_.Lock();
   }
@@ -1129,15 +1152,11 @@ Status DBImpl::WriteLevel0Table(std::vector<MemTable*> &mems, VersionEdit* edit,
     Log(options_.info_log,
         "Level-0 flush table #%lu: started",
         (unsigned long)meta.number);
-    // We skip compression if universal compression is used and the size
-    // threshold is set for compression.
-    bool enable_compression = (options_.compaction_style
-        != kCompactionStyleUniversal ||
-        options_.compaction_options_universal.compression_size_percent < 0);
+
     s = BuildTable(dbname_, env_, options_, storage_options_,
                    table_cache_.get(), iter, &meta,
                    user_comparator(), newest_snapshot,
-                   earliest_seqno_in_memtable, enable_compression);
+                   earliest_seqno_in_memtable, GetCompressionFlush(options_));
     LogFlush(options_.info_log);
     delete iter;
     Log(options_.info_log, "Level-0 flush table #%lu: %lu bytes %s",
diff --git a/db/db_impl.h b/db/db_impl.h
index adeb163f0..d33efd19e 100644
--- a/db/db_impl.h
+++ b/db/db_impl.h
@@ -589,4 +589,7 @@ extern Options SanitizeOptions(const std::string& db,
 CompressionType GetCompressionType(const Options& options, int level,
                                    const bool enable_compression);
 
+// Determine compression type for L0 file written by memtable flush.
+CompressionType GetCompressionFlush(const Options& options);
+
 }  // namespace rocksdb
diff --git a/db/repair.cc b/db/repair.cc
index fc9ba282d..6db90c865 100644
--- a/db/repair.cc
+++ b/db/repair.cc
@@ -225,7 +225,8 @@ class Repairer {
     Iterator* iter = mem->NewIterator();
     status = BuildTable(dbname_, env_, options_, storage_options_,
                         table_cache_, iter, &meta,
-                        icmp_.user_comparator(), 0, 0, true);
+                        icmp_.user_comparator(), 0, 0,
+                        kNoCompression);
     delete iter;
     delete mem->Unref();
     mem = nullptr;