SERVER-32970 put back random chunk size initialization in mongos auto split tracking

3 files changed, 27 insertions, 6 deletions

diff --git a/jstests/sharding/write_cmd_auto_split.js b/jstests/sharding/write_cmd_auto_split.js
index ab5ca26c84f..74155d8b0bf 100644
--- a/jstests/sharding/write_cmd_auto_split.js
+++ b/jstests/sharding/write_cmd_auto_split.js
@@ -40,7 +40,7 @@
 
     assert.eq(1, configDB.chunks.find({"ns": "test.update"}).itcount());
 
-    for (var x = 0; x < 1100; x++) {
+    for (var x = 0; x < 2100; x++) {
         assert.writeOK(testDB.runCommand({
             update: 'update',
             updates: [{q: {x: x}, u: {x: x, v: doc1k}, upsert: true}],
@@ -80,7 +80,7 @@
 
     // Note: Estimated 'chunk size' tracked by mongos is initialized with a random value so
     // we are going to be conservative.
-    for (var x = 0; x < 1100; x += 400) {
+    for (var x = 0; x < 2100; x += 400) {
         var docs = [];
 
         for (var y = 0; y < 400; y++) {
@@ -101,7 +101,7 @@
 
     assert.eq(1, configDB.chunks.find({"ns": "test.update"}).itcount());
 
-    for (var x = 0; x < 1100; x += 400) {
+    for (var x = 0; x < 2100; x += 400) {
         var docs = [];
 
         for (var y = 0; y < 400; y++) {
@@ -123,7 +123,7 @@
 
     assert.eq(1, configDB.chunks.find({"ns": "test.delete"}).itcount());
 
-    for (var x = 0; x < 1100; x += 400) {
+    for (var x = 0; x < 2100; x += 400) {
         var docs = [];
 
         for (var y = 0; y < 400; y++) {
diff --git a/src/mongo/s/chunk.cpp b/src/mongo/s/chunk.cpp
index 3a25fa1c055..527440594eb 100644
--- a/src/mongo/s/chunk.cpp
+++ b/src/mongo/s/chunk.cpp
@@ -32,16 +32,37 @@
 
 #include "mongo/s/chunk.h"
 
+#include "mongo/platform/random.h"
+#include "mongo/s/grid.h"
 #include "mongo/util/mongoutils/str.h"
 
 namespace mongo {
 
+namespace {
+
+// Test whether we should split once data * kSplitCheckInterval > chunkSize (approximately)
+PseudoRandom prng(static_cast<int64_t>(time(0)));
+
+// Assume user has 64MB chunkSize setting. It is ok if this assumption is wrong since it is only
+// a heuristic.
+const int64_t kMaxDataWritten = 64 / Chunk::kSplitTestFactor;
+
+/**
+ * Generates a random value for _dataWritten so that a mongos restart wouldn't cause delay in
+ * splitting.
+ */
+int64_t mkDataWritten() {
+    return prng.nextInt64(kMaxDataWritten);
+}
+
+}  // namespace
+
 Chunk::Chunk(const ChunkType& from)
     : _range(from.getMin(), from.getMax()),
       _shardId(from.getShard()),
       _lastmod(from.getVersion()),
       _jumbo(from.getJumbo()),
-      _dataWritten(0) {
+      _dataWritten(mkDataWritten()) {
     invariantOK(from.validate());
 }
 
diff --git a/src/mongo/s/chunk.h b/src/mongo/s/chunk.h
index 867bdce54fc..540238463b5 100644
--- a/src/mongo/s/chunk.h
+++ b/src/mongo/s/chunk.h
@@ -42,7 +42,7 @@ class BSONObj;
 class Chunk {
 public:
     // Test whether we should split once data * kSplitTestFactor > chunkSize (approximately)
-    const uint64_t kSplitTestFactor = 5;
+    static const uint64_t kSplitTestFactor = 5;
 
     explicit Chunk(const ChunkType& from);