fix: Haskell OOM across datasets

_benchplot/benchplot-dict.R

@@ -369,7 +369,7 @@ groupby.data.exceptions = {list(
   "chdb" = {list(
   )},
   "haskell" = {list(
-  	"Not tested/OOM" = c("G1_1e8_1e2_0_0", "G1_1e8_1e1_0_0", "G1_1e8_2e0_0_0", "G1_1e8_1e2_0_1", "G1_1e8_1e2_5_0", "G1_1e9_1e2_0_0","G1_1e9_1e2_0_1","G1_1e9_1e2_5_0","G1_1e9_1e1_0_0", "G1_1e9_2e0_0_0")
+  	"Not tested/OOM" = c("G1_1e8_1e2_5_0", "G1_1e9_1e2_0_0","G1_1e9_1e2_0_1","G1_1e9_1e2_5_0","G1_1e9_1e1_0_0", "G1_1e9_2e0_0_0")
   )}
 )}
 groupby.exceptions = task.exceptions(groupby.query.exceptions, groupby.data.exceptions)

haskell/VERSION

@@ -1 +1 @@
-2.2.0.0
+3.6.0.0

haskell/cabal.project

@@ -1,2 +1,3 @@
 packages: .
 allow-newer: all
+optimization: 2

haskell/exec.sh

@@ -24,4 +24,23 @@ fi
 source ./haskell/ver-haskell.sh
 
 [ -d "$HOME/.cache/cabal/packages/hackage.haskell.org" ] || cabal update
+
+## Cap GHC's heap at 90% of machine memory so we exit gracefully.
+if [ -r /proc/meminfo ]; then
+    MEM_KB=$(awk '/^MemTotal:/ {print $2}' /proc/meminfo)
+else
+    MEM_KB=$(( $(sysctl -n hw.memsize) / 1024 ))
+fi
+export GHCRTS="-M$(( MEM_KB * 9 / 10 / 1024 ))m"
+
+## Best-effort THP=always (GHC heap on 2MB pages); skipped when sudo is unavailable.
+THP=/sys/kernel/mm/transparent_hugepage
+if [ -w "$THP/enabled" ]; then
+    echo always > "$THP/enabled" || true
+    echo defer+madvise > "$THP/defrag" || true
+elif sudo -n true 2>/dev/null; then
+    echo always | sudo tee "$THP/enabled" > /dev/null || true
+    echo defer+madvise | sudo tee "$THP/defrag" > /dev/null || true
+fi
+
 cabal --project-dir=./haskell run -O2 "$1-haskell"

haskell/groupby-haskell.hs

@@ -6,11 +6,9 @@ import BenchmarkCommon
 import Control.Arrow ((>>>))
 import Control.Monad (forM_)
 import Data.Functor ((<&>))
-import qualified Data.List as L
 import Data.Maybe (fromMaybe)
 import Data.Text (Text)
 import qualified Data.Text as T
-import qualified Data.Vector as V
 import qualified Data.Vector.Unboxed as VU
 import DataFrame.Functions ((.=))
 import qualified DataFrame.Functions as F
@@ -21,12 +19,14 @@ import DataFrame.IO.CSV (
  )
 import DataFrame.IO.CSV.Fast
 import DataFrame.Internal.DataFrame (DataFrame)
-import DataFrame.Internal.Expression (AggStrategy (..), Expr (..))
-import DataFrame.Schema (schemaType)
+import DataFrame.Internal.Expression (Expr (..))
 import qualified DataFrame.Operations.Aggregation as D
 import qualified DataFrame.Operations.Core as D
+-- Semigroup DataFrame (row concat)
+import DataFrame.Operations.Merge ()
 import qualified DataFrame.Operations.Subset as D
 import qualified DataFrame.Operations.Transformations as D
+import DataFrame.Schema (schemaType)
 import System.Environment (getEnv, lookupEnv)
 import System.IO (BufferMode (..), hPutStrLn, hSetBuffering, stderr, stdout)
 
@@ -111,7 +111,7 @@ runBenchmark srcFile dataName machineType = do
         config
         df
         "sum v1 by id1"
-        (D.groupBy [F.name id1] >>> D.aggregate [F.sum v1 `F.as` "v1_sum"])
+        (D.groupBy [F.name id1] >>> D.aggregate ["v1_sum" .= F.sum v1])
         (\res -> [chkSumInt "v1_sum" res])
 
     -- Q2: Sum v1 by id1:id2
@@ -174,36 +174,43 @@ runBenchmark srcFile dataName machineType = do
         (D.groupBy [F.name id3] >>> D.aggregate ["diff" .= F.maximum v1 - F.minimum v2])
         (\res -> [chkSumInt "diff" res])
 
-    -- Q8: Largest two v3 by id6.
-    -- Polars takes top_k(2) per group then explodes and sums; the sum of the
-    -- exploded values equals the sum of (max + 2nd-max) per group, which is
-    -- what 'top2Sum' computes per group. Checksum parity is preserved.
+    -- Q8: largest two v3 by id6, exploded to two rows per group
+    -- (polars reference shape: group_by(id6).agg(v3.top_k(2)).explode(v3)).
+    -- 'F.topK' returns the pair in descending order; the library has no explode,
+    -- so we take the two ends and stack the halves.
     runQuestion
         config
         df
         "largest two v3 by id6"
-        (D.groupBy [F.name id6] >>> D.aggregate ["largest2_v3" .= top2Sum v3])
-        (\res -> [chkSumDbl "largest2_v3" res])
+        ( \d ->
+            let top2 = F.col @[Double] "top2"
+                r =
+                    ( D.groupBy [F.name id6]
+                        >>> D.aggregate ["top2" .= F.topK 2 v3]
+                        >>> D.derive "v3" (F.fromMaybe nan (F.firstOrNothing top2))
+                        >>> D.derive "v3b" (F.fromMaybe nan (F.lastOrNothing top2))
+                    )
+                        d
+             in D.select ["id6", "v3"] r
+                    <> D.rename "v3b" "v3" (D.select ["id6", "v3b"] r)
+        )
+        (\res -> [chkSumDbl "v3" res])
 
     -- Q9: Regression (r^2 of v1 vs v2) by id2, id4.
     -- corr(v1,v2)^2 from the per-group sums of v1, v2, v1*v2, v1^2, v2^2 and the
     -- group count, then summed. Matches polars pl.corr(...)**2 then .sum().
     runQuestion
         config
-        ( D.derive
-            "v2v2"
-            (dv2 * dv2)
-            (D.derive "v1v1" (dv1 * dv1) (D.derive "v1v2" (dv1 * dv2) df))
-        )
+        df
         "regression v1 v2 by id2 id4"
         ( D.groupBy [F.name id2, F.name id4]
             >>> D.aggregate
                 [ "n" .= F.count v1
                 , "sx" .= F.sum dv1
                 , "sy" .= F.sum dv2
-                , "sxy" .= F.sum (F.col @Double "v1v2")
-                , "sxx" .= F.sum (F.col @Double "v1v1")
-                , "syy" .= F.sum (F.col @Double "v2v2")
+                , "sxy" .= F.sum (dv1 * dv2)
+                , "sxx" .= F.sum (dv1 * dv1)
+                , "syy" .= F.sum (dv2 * dv2)
                 ]
             >>> D.derive "r2" r2Expr
             >>> D.select [F.name id2, F.name id4, "r2"]
@@ -244,6 +251,12 @@ runQuestion cfg inputDF qLabel transform chkFn = do
         writeLog cfg qLabel outRows outCols runNum calcTime memUsage chkValues chkTime
     putStrLn $ qLabel ++ " completed."
 
+{- | Stand-in for a group with fewer than two values. The benchmark data has no
+size-1 id6 groups, so this never reaches the checksum.
+-}
+nan :: Double
+nan = 0 / 0
+
 chkSumInt :: String -> DataFrame -> Double
 chkSumInt col df =
     case D.columnAsIntVector (F.col @Int (T.pack col)) df of
@@ -256,13 +269,6 @@ chkSumDbl col df =
         Right vec -> VU.sum vec
         Left _ -> 0.0
 
--- | Per-group sum of the two largest values (top_k(2) then sum, like polars Q8).
-top2Sum :: Expr Double -> Expr Double
-top2Sum = Agg (CollectAgg "top2Sum" f)
-  where
-    f :: V.Vector Double -> Double
-    f v = Prelude.sum (take 2 (L.sortBy (flip compare) (V.toList v)))
-
 {- | Squared Pearson correlation of v1,v2 from per-group moment sums (polars Q9).
 Degenerate groups (zero variance) contribute 0 rather than poisoning the sum
 with NaN; the 1e7 data has no such groups so the checksum is unaffected.

haskell/haskell-benchmark.cabal

@@ -7,19 +7,19 @@ common df-deps
   build-depends:       base >= 4.7 && < 5
                      , bytestring >= 0.11 && < 0.13
                      , cassava >= 0.1 && < 1
-                     , dataframe-core ^>= 2.1
-                     , dataframe-parsing ^>= 2.1
-                     , dataframe-csv ^>= 2.1
-                     , dataframe-operations ^>= 2.1
-                     , dataframe-fastcsv ^>= 1.3
+                     , dataframe-core ^>= 2.5
+                     , dataframe-parsing ^>= 2.2
+                     , dataframe-csv ^>= 2.3
+                     , dataframe-operations ^>= 2.5
+                     , dataframe-fastcsv ^>= 1.4.1
                      , text >= 2.1
                      , vector >= 0.12
                      , time >= 1.9
                      , process >= 1.6
                      , directory >= 1.3
                      , unix
   default-language:    Haskell2010
-  ghc-options:         -O2 -threaded -rtsopts "-with-rtsopts=-N8 -c -F1.5 -M30g -Fd1"
+  ghc-options:         -O2 -threaded -rtsopts "-with-rtsopts=-N16 -A64m -n4m -I0 -c -F1.5 -M30g -Fd1"
 
 executable groupby-haskell
   import:              df-deps
@@ -32,9 +32,3 @@ executable join-haskell
   main-is:             join-haskell.hs
   hs-source-dirs:      .
   other-modules:       BenchmarkCommon
-
-executable timeq
-  import:              df-deps
-  main-is:             timeq.hs
-  hs-source-dirs:      .
-  build-depends:       deepseq

haskell/join-haskell.hs

@@ -19,13 +19,14 @@ import DataFrame.IO.CSV (
     defaultReadOptions,
  )
 import DataFrame.IO.CSV.Fast (fastReadCsvWithOpts)
-import DataFrame.Internal.DataFrame (DataFrame)
-import DataFrame.Schema (SchemaType, schemaType)
+import DataFrame.Internal.DataFrame (DataFrame, forceDataFrame)
 import qualified DataFrame.Operations.Core as D
 import qualified DataFrame.Operations.Join as DJ
+import DataFrame.Schema (SchemaType, schemaType)
 import Numeric (showEFloat)
 import System.Environment (getEnv, lookupEnv)
 import System.IO (BufferMode (..), hSetBuffering, stdout)
+import System.Mem (performGC)
 
 main :: IO ()
 main = do
@@ -148,10 +149,9 @@ runJoin ::
     IO ()
 runJoin cfg leftDF rightDF qLabel joinFn = do
     forM_ [1, 2] $ \runNum -> do
+        performGC
         (resultDF, calcTime) <- timeIt $ do
-            let res = freshRun runNum (uncurry joinFn) (leftDF, rightDF)
-            _ <- evaluate res
-            return res
+            evaluate (forceDataFrame (freshRun runNum (uncurry joinFn) (leftDF, rightDF)))
 
         memUsage <- getMemoryUsage
         let (outRows, outCols) = D.dimensions resultDF

haskell/ver-haskell.sh

@@ -3,7 +3,7 @@ set -e
 
 cd haskell
 
-export HASKELL_DF_VERSION="2.2.0.0"
+export HASKELL_DF_VERSION="3.6.0.0"
 
 echo "${HASKELL_DF_VERSION}" > VERSION
 

logs.csv

@@ -7392,23 +7392,3 @@ ip-172-31-18-189,1784046154,duckdb,1.5.4,08e34c447b,groupby,G1_1e9_1e2_5_0,17840
 ip-172-31-18-189,1784046154,duckdb,1.5.4,08e34c447b,groupby,G1_1e9_1e2_5_0,1784053271.16982,finish,0,0,c6id.4xlarge
 ip-172-31-18-189,1784046154,duckdb,1.5.4,08e34c447b,join,J1_1e9_NA_0_0,1784053286.1849,start,,,c6id.4xlarge
 ip-172-31-18-189,1784046154,duckdb,1.5.4,08e34c447b,join,J1_1e9_NA_0_0,1784055922.71045,finish,0,0,c6id.4xlarge
-ip-172-31-18-189,1786373405,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e7_1e2_0_0,1786373407.63123,start,,,c6id.4xlarge
-ip-172-31-18-189,1786373405,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e7_1e2_0_0,1786373435.22305,finish,7,0,c6id.4xlarge
-ip-172-31-18-189,1786373405,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e7_1e1_0_0,1786373450.23815,start,,,c6id.4xlarge
-ip-172-31-18-189,1786373405,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e7_1e1_0_0,1786373477.74633,finish,7,0,c6id.4xlarge
-ip-172-31-18-189,1786373405,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e7_2e0_0_0,1786373492.76142,start,,,c6id.4xlarge
-ip-172-31-18-189,1786373405,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e7_2e0_0_0,1786373527.4555,finish,11,0,c6id.4xlarge
-ip-172-31-18-189,1786373405,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e7_1e2_0_1,1786373542.47061,start,,,c6id.4xlarge
-ip-172-31-18-189,1786373405,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e7_1e2_0_1,1786373567.22655,finish,7,0,c6id.4xlarge
-ip-172-31-18-189,1786373405,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e7_1e2_5_0,1786373582.24164,start,,,c6id.4xlarge
-ip-172-31-18-189,1786373405,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e7_1e2_5_0,1786373582.84261,finish,8,0,c6id.4xlarge
-ip-172-31-18-189,1786373702,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e8_1e2_0_0,1786373703.5071,start,,,c6id.4xlarge
-ip-172-31-18-189,1786373702,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e8_1e2_0_0,1786373734.20329,finish,10,251,c6id.4xlarge
-ip-172-31-18-189,1786373702,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e8_1e1_0_0,1786373749.21738,start,,,c6id.4xlarge
-ip-172-31-18-189,1786373702,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e8_1e1_0_0,1786373779.22521,finish,10,251,c6id.4xlarge
-ip-172-31-18-189,1786373702,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e8_2e0_0_0,1786373794.24029,start,,,c6id.4xlarge
-ip-172-31-18-189,1786373702,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e8_2e0_0_0,1786373824.67766,finish,14,251,c6id.4xlarge
-ip-172-31-18-189,1786373702,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e8_1e2_0_1,1786373839.69275,start,,,c6id.4xlarge
-ip-172-31-18-189,1786373702,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e8_1e2_0_1,1786373869.93667,finish,10,251,c6id.4xlarge
-ip-172-31-18-189,1786373702,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e8_1e2_5_0,1786373884.95176,start,,,c6id.4xlarge
-ip-172-31-18-189,1786373702,haskell,2.2.0.0,dataframe-2.2.0.0,groupby,G1_1e8_1e2_5_0,1786373885.54634,finish,8,0,c6id.4xlarge