Grow TxQ expected size quickly, shrink slowly (RIPD-1534):

* Stores recent history of "good" ledgers. Uses the maximum as the
  expected ledger size. When a large value drops off, use a 90%
  backoff to go down to to the new maximum.
* If consensus is unhealthy, wipe the history in addition to the current
  clamping.
* Include .md doc files in xcode and VS projects

CMakeLists.txt

@@ -337,7 +337,7 @@ list(APPEND non_unity_srcs "${test_srcs}")
 
 if(WIN32 OR is_xcode)
     # Rippled headers. Only needed for IDEs.
-    file(GLOB_RECURSE rippled_headers src/*.h src/*.hpp)
+    file(GLOB_RECURSE rippled_headers src/*.h src/*.hpp *.md)
     list(APPEND rippled_headers Builds/CMake/CMakeFuncs.cmake)
     foreach(curdir
             beast/asio

src/ripple/app/misc/FeeEscalation.md

@@ -28,8 +28,12 @@ consensus process, but will be at least [5](#other-constants).
   * If consensus stays [healthy](#consensus-health), the limit will
   be the max of the current limit or the number of transactions in
   the validated ledger until it gets to [50](#other-constants), at
-  which point, the limit will only be updated to the number of
-  transactions in the validated ledger if it is larger than 50.
+  which point, the limit will be the largest number of transactions
+  in the last [20](#other-constants) validated ledgers which had
+  more than 50 transactions. Any time the limit decreases (ie. a
+  large ledger is no longer recent), the limit will decrease to the
+  new largest value by 10% each time the ledger has more than 50
+  transactions.
   * If consensus does not stay [healthy](#consensus-health),
   the limit will clamp down to the smaller of [50](#other-constants)
   or the number of transactions in the validated ledger.
@@ -109,7 +113,8 @@ but in practice, either
   sequence number and at least a [25% higher fee](#other-constants), or
   * it will get dropped when the queue fills up with more valuable transactions.
   The size limit is computed dynamically, and can hold transactions for
-  the next [20 ledgers](#other-constants). The lower the transaction's
+  the next [20 ledgers](#other-constants) (restricted to a minimum of
+  [2000 transactions](#other-constants)). The lower the transaction's
   fee, the more likely that it will get dropped if the network is busy.
 
 If a transaction is submitted for an account with one or more transactions
@@ -200,7 +205,9 @@ automatically as the ripple network's performance improves, allowing
 more transactions per second, and thus more transactions per ledger
 to process successfully.  The limit of 20 ledgers was used to provide
 a balance between resource (specifically memory) usage, and giving
-transactions a realistic chance to be processed. This exact value was
+transactions a realistic chance to be processed. The minimum size of
+2000 transactions was chosen to allow a decent functional backlog during
+network congestion conditions. These exact values were
 chosen experimentally, and can easily change in the future.
 * *Maximum retries*. A transaction in the queue can attempt to apply
 to the open ledger, but get a retry (`ter`) code up to 10 times, at

src/ripple/app/misc/NetworkOPs.cpp

@@ -1099,7 +1099,7 @@ void NetworkOPsImp::apply (std::unique_lock<std::mutex>& batchLock)
             }
             else if (e.result == terQUEUED)
             {
-                JLOG(m_journal.info()) << "Transaction is likely to claim a " <<
+                JLOG(m_journal.debug()) << "Transaction is likely to claim a " <<
                     "fee, but is queued until fee drops";
                 e.transaction->setStatus(HELD);
                 // Add to held transactions, because it could get

src/ripple/app/misc/TxQ.h

@@ -26,6 +26,7 @@
 #include <ripple/protocol/TER.h>
 #include <ripple/protocol/STTx.h>
 #include <boost/intrusive/set.hpp>
+#include <boost/circular_buffer.hpp>
 
 namespace ripple {
 
@@ -194,6 +195,9 @@ class TxQ
         // One more than this value will be accepted
         // before escalation kicks in.
         std::size_t txnsExpected_;
+        // Recent history of transaction counts that
+        // exceed the targetTxnCount_
+        boost::circular_buffer<std::size_t> recentTxnCounts_;
         // Minimum value of escalationMultiplier.
         std::uint64_t const minimumMultiplier_;
         // Based on the median fee of the LCL. Used
@@ -213,6 +217,7 @@ class TxQ
                     targetTxnCount_ : *setup.maximumTxnInLedger :
                         boost::optional<std::size_t>(boost::none))
             , txnsExpected_(minimumTxnCount_)
+            , recentTxnCounts_(setup.ledgersInQueue)
             , minimumMultiplier_(setup.minimumEscalationMultiplier)
             , escalationMultiplier_(minimumMultiplier_)
             , j_(j)

src/ripple/app/misc/impl/TxQ.cpp

@@ -29,6 +29,7 @@
 #include <boost/algorithm/clamp.hpp>
 #include <limits>
 #include <numeric>
+#include <algorithm>
 
 namespace ripple {
 
@@ -107,16 +108,32 @@ TxQ::FeeMetrics::update(Application& app,
         // so clamp down on limits.
         txnsExpected_ = boost::algorithm::clamp(feeLevels.size(),
             minimumTxnCount_, targetTxnCount_);
+        recentTxnCounts_.clear();
     }
     else if (feeLevels.size() > txnsExpected_ ||
         feeLevels.size() > targetTxnCount_)
     {
+        recentTxnCounts_.push_back(feeLevels.size());
+        auto const iter = std::max_element(recentTxnCounts_.begin(),
+            recentTxnCounts_.end());
+        BOOST_ASSERT(iter != recentTxnCounts_.end());
+        auto const next = [&]
+        {
+            // Grow quickly: If the max_element is >= the
+            // current size limit, use it.
+            if (*iter >= txnsExpected_)
+                return *iter;
+            // Shrink slowly: If the max_element is < the
+            // current size limit, use a limit that is
+            // 90% of the way from max_element to the
+            // current size limit.
+            return (txnsExpected_ * 9 + *iter) / 10;
+        }();
         // Ledgers are processing in a timely manner,
         // so keep the limit high, but don't let it
         // grow without bound.
-        txnsExpected_ = maximumTxnCount_ ?
-            std::min(feeLevels.size(), *maximumTxnCount_) :
-            feeLevels.size();
+        txnsExpected_ = std::min(next,
+            maximumTxnCount_.value_or(next));
     }
 
     if (feeLevels.empty())