From a6e0932885e03b6c1c8007d76b619f86f1f09c7e Mon Sep 17 00:00:00 2001
From: Benjamin Wagner <benjamin.wagner@firebolt.io>
Date: Mon, 6 Nov 2023 15:51:22 +0100
Subject: [PATCH] Add Big Join Query

Based on Peter's advice to see if my microbenchmarks generalize to a
very simple query - turns out they do!

We add `q_bigjoin`: `SELECT count(*) FROM lineitem, orders WHERE
l_orderkey = o_orderkey`.

For this query at SF 10 we can actually nicely see that Interpretation &
ROF start winning against JIT compilation because they can issue more
independent hash table loads.
---
 src/common/TPCH.cpp        | 53 ++++++++++++++++++++++++++++++++++++++
 src/common/TPCH.h          |  1 +
 test/tpch/test_queries.cpp |  4 ++-
 tools/inkfuse_bench.cpp    |  1 +
 tools/inkfuse_runner.cpp   |  3 +++
 5 files changed, 61 insertions(+), 1 deletion(-)

diff --git a/src/common/TPCH.cpp b/src/common/TPCH.cpp
index 07da0f8..c645bf2 100644
--- a/src/common/TPCH.cpp
+++ b/src/common/TPCH.cpp
@@ -1784,6 +1784,59 @@ std::unique_ptr<Print> q19(const Schema& schema) {
 
 }
 
+std::unique_ptr<Print> q_bigjoin(const inkfuse::Schema& schema) {
+   // Join orders onto lineitem.
+   auto& rel_o = schema.at("orders");
+   std::vector<std::string> cols_o{
+      "o_orderkey"};
+   auto scan_o = TableScan::build(*rel_o, cols_o, "scan_o");
+   auto& scan_ref_o = *scan_o;
+
+   auto& rel_l = schema.at("lineitem");
+   std::vector<std::string> cols_l{
+      "l_orderkey"};
+   auto scan_l = TableScan::build(*rel_l, cols_l, "scan_l");
+   auto& scan_ref_l = *scan_l;
+
+   std::vector<RelAlgOpPtr> o_l_join_children;
+   o_l_join_children.push_back(std::move(scan_o));
+   o_l_join_children.push_back(std::move(scan_l));
+   auto o_l_join = Join::build(
+      std::move(o_l_join_children),
+      "o_l_join",
+      // Keys left (o_orderkey)
+      {scan_ref_o.getOutput()[0]},
+      // Payload left ()
+      {},
+      // Keys right (l_orderkey)
+      {scan_ref_l.getOutput()[0]},
+      {},
+      JoinType::Inner,
+      true);
+   auto& o_l_join_ref = *o_l_join;
+
+   // 2. Aggregate count(*).
+   std::vector<RelAlgOpPtr> agg_children;
+   agg_children.push_back(std::move(o_l_join));
+   // Don't group by anything on this query.
+   std::vector<const IU*> group_by{};
+   std::vector<AggregateFunctions::Description> aggregates{
+      {*o_l_join_ref.getOutput()[0], AggregateFunctions::Opcode::Count}};
+   auto agg = Aggregation::build(
+      std::move(agg_children),
+      "agg",
+      std::move(group_by),
+      std::move(aggregates));
+
+   // 6. Print
+   std::vector<const IU*> out_ius{agg->getOutput()[0]};
+   std::vector<std::string> colnames = {"num_rows"};
+   std::vector<RelAlgOpPtr> print_children;
+   print_children.push_back(std::move(agg));
+   return Print::build(std::move(print_children),
+                       std::move(out_ius), std::move(colnames));
+}
+
 std::unique_ptr<Print> l_count(const inkfuse::Schema& schema) {
    // 1. Scan from lineitem.
    auto& rel = schema.at("lineitem");
diff --git a/src/common/TPCH.h b/src/common/TPCH.h
index 0df550a..dd7da71 100644
--- a/src/common/TPCH.h
+++ b/src/common/TPCH.h
@@ -34,6 +34,7 @@ std::unique_ptr<Print> q18(const Schema& schema);
 std::unique_ptr<Print> q19(const Schema& schema);
 
 /// Some interesting custom queries. See /tpch for query text.
+std::unique_ptr<Print> q_bigjoin(const Schema& schema);
 std::unique_ptr<Print> l_count(const Schema& schema);
 std::unique_ptr<Print> l_point(const Schema& schema);
 
diff --git a/test/tpch/test_queries.cpp b/test/tpch/test_queries.cpp
index 3abd1e2..1a3f112 100644
--- a/test/tpch/test_queries.cpp
+++ b/test/tpch/test_queries.cpp
@@ -38,6 +38,7 @@ const std::unordered_map<std::string, FunctionT> generator_map{
    {"q14", tpch::q14},
    {"q18", tpch::q18},
    {"q19", tpch::q19},
+   {"q_bigjoin", tpch::q_bigjoin},
    {"l_count", tpch::l_count},
    {"l_point", tpch::l_point},
 };
@@ -53,6 +54,7 @@ std::unordered_map<std::string, size_t> expected_rows{
    {"q14", 1},
    {"q18", 0},
    {"q19", 0},
+   {"q_bigjoin", 1},
    {"l_count", 1},
    {"l_point", 6},
 };
@@ -76,7 +78,7 @@ INSTANTIATE_TEST_CASE_P(
    tpch_queries,
    TPCHQueriesTestT,
    ::testing::Combine(
-      ::testing::Values("q1", "q3", "q4", "q5", "q6", "q14", "q18", "q19", "l_count", "l_point"),
+      ::testing::Values("q1", "q3", "q4", "q5", "q6", "q14", "q18", "q19", "l_count", "q_bigjoin", "l_point"),
       ::testing::Values(
          PipelineExecutor::ExecutionMode::Fused,
          PipelineExecutor::ExecutionMode::Interpreted,
diff --git a/tools/inkfuse_bench.cpp b/tools/inkfuse_bench.cpp
index e4c58d4..06d625d 100644
--- a/tools/inkfuse_bench.cpp
+++ b/tools/inkfuse_bench.cpp
@@ -43,6 +43,7 @@ const std::vector<std::pair<std::string, decltype(tpch::q1)*>> queries = {
    {"q14", tpch::q14},
    {"q18", tpch::q18},
    {"q19", tpch::q19},
+   {"q_bigjoin", tpch::q_bigjoin},
    {"l_count", tpch::l_count},
    {"l_point", tpch::l_point},
 };
diff --git a/tools/inkfuse_runner.cpp b/tools/inkfuse_runner.cpp
index f89af5a..cbdbd86 100644
--- a/tools/inkfuse_runner.cpp
+++ b/tools/inkfuse_runner.cpp
@@ -198,6 +198,9 @@ int main(int argc, char* argv[]) {
                } else if (split[1] == "q19") {
                   auto q = tpch::q19(*loaded);
                   runQuery("q19", std::move(q), mode, thread_count);
+               } else if (split[1] == "q_bigjoin") {
+                  auto q = tpch::q_bigjoin(*loaded);
+                  runQuery("q_bigjoin", std::move(q), mode, thread_count);
                } else if (split[1] == "l_count") {
                   auto q = tpch::l_count(*loaded);
                   runQuery("l_count", std::move(q), mode, thread_count);