Introduce the "report" command to combine multiple evaluations into a…

… single file

Closes #205

cmd/eval-dev-quality/cmd/command.go

@@ -9,6 +9,7 @@ import (
 type Command struct {
 	Evaluate     `command:"evaluate" description:"Run an evaluation, by default with all defined models, repositories and tasks."`
 	InstallTools `command:"install-tools" description:"Checks and installs all tools required for the evaluation benchmark."`
+	Report       `command:"report" description:"Combines the results of multiple evaluations."`
 	Version      `command:"version" description:"Display the version information of the binary."`
 }
 

cmd/eval-dev-quality/cmd/command_test.go

@@ -35,7 +35,7 @@ func TestExecute(t *testing.T) {
 
 			ExpectedOutput: bytesutil.StringTrimIndentations(`
 				Usage:
-				  eval-dev-quality [OPTIONS] [evaluate | install-tools | version]
+				  eval-dev-quality [OPTIONS] [command]
 
 				Command to manage, update and actually execute the ` + "`" + `eval-dev-quality` + "`" + `
 				evaluation benchmark.
@@ -47,6 +47,7 @@ func TestExecute(t *testing.T) {
 				Available commands:
 				  evaluate       Run an evaluation, by default with all defined models, repositories and tasks.
 				  install-tools  Checks and installs all tools required for the evaluation benchmark.
+				  report         Combines the results of multiple evaluations.
 				  version        Display the version information of the binary.
 			`),
 		})
@@ -56,7 +57,7 @@ func TestExecute(t *testing.T) {
 
 			ExpectedOutput: bytesutil.StringTrimIndentations(`
 				Usage:
-				  eval-dev-quality [OPTIONS] [evaluate | install-tools | version]
+				  eval-dev-quality [OPTIONS] [command]
 
 				Command to manage, update and actually execute the ` + "`" + `eval-dev-quality` + "`" + `
 				evaluation benchmark.
@@ -67,6 +68,7 @@ func TestExecute(t *testing.T) {
 				Available commands:
 				  evaluate       Run an evaluation, by default with all defined models, repositories and tasks.
 				  install-tools  Checks and installs all tools required for the evaluation benchmark.
+				  report         Combines the results of multiple evaluations.
 				  version        Display the version information of the binary.
 			`),
 		})

cmd/eval-dev-quality/cmd/report.go

@@ -0,0 +1,107 @@
+package cmd
+
+import (
+	"os"
+	"path/filepath"
+	"sort"
+
+	pkgerrors "github.com/pkg/errors"
+	"golang.org/x/exp/maps"
+
+	"github.com/symflower/eval-dev-quality/evaluate/report"
+	"github.com/symflower/eval-dev-quality/log"
+)
+
+// Report holds the "report" command.
+type Report struct {
+	// EvaluationPaths holds the list of file paths where the results of previous evaluations are stored.
+	EvaluationPaths []string `long:"evaluation-path" description:"File path for an evaluation CSV file where the results of a previous evaluation are stored."`
+	// ResultPath holds the directory path where the overall results should be written to.
+	ResultPath string `long:"result-path" description:"Directory path were the combined results are written to."`
+
+	// logger holds the logger of the command.
+	logger *log.Logger
+}
+
+var _ SetLogger = (*Evaluate)(nil)
+
+// SetLogger sets the logger of the command.
+func (command *Report) SetLogger(logger *log.Logger) {
+	command.logger = logger
+}
+
+// Execute executes the command.
+func (command *Report) Execute(args []string) (err error) {
+	// Create the result path directory and check if there is already an evaluation CSV file in there.
+	var evaluationCSVFile *os.File
+	if err = os.MkdirAll(filepath.Dir(command.ResultPath), 0755); err != nil {
+		command.logger.Panicf("ERROR: %s", err)
+	}
+	if _, err := os.Stat(filepath.Join(command.ResultPath, "evaluation.csv")); err != nil {
+		if os.IsNotExist(err) {
+			evaluationCSVFile, err = os.Create(filepath.Join(command.ResultPath, "evaluation.csv"))
+			if err != nil {
+				command.logger.Panicf("ERROR: %s", err)
+			}
+			defer evaluationCSVFile.Close()
+		} else {
+			command.logger.Panicf("ERROR: %s", err)
+		}
+	} else {
+		command.logger.Panicf("ERROR: an evaluation CSV file already exists in %s", command.ResultPath)
+	}
+
+	var allRecords []string
+	for _, evaluationPath := range command.EvaluationPaths {
+		// Collect all evaluation CSV file paths.
+		evaluationFilePaths, err := pathsFromGlobPattern(evaluationPath)
+		if err != nil {
+			command.logger.Panicf("ERROR: %s", err)
+		}
+		// Collect all records from the evaluation CSV files.
+		records, err := report.RecordsFromEvaluationCSVFiles(evaluationFilePaths)
+		if err != nil {
+			command.logger.Panicf("ERROR: %s", err)
+		}
+		allRecords = append(allRecords, records...)
+	}
+
+	if len(allRecords) == 0 {
+		command.logger.Printf("no evaluation records found in %+v", command.EvaluationPaths)
+
+		return nil
+	}
+	sort.Strings(allRecords)
+
+	// Write all records into a single evaluation CSV file.
+	evaluationFile, err := report.NewEvaluationFile(evaluationCSVFile)
+	if err != nil {
+		command.logger.Panicf("ERROR: %s", err)
+	}
+	if err := evaluationFile.WriteEvaluationRawRecords(allRecords); err != nil {
+		command.logger.Panicf("ERROR: %s", err)
+	}
+
+	return nil
+}
+
+// pathsFromGlobPattern returns all evaluation CSV file paths.
+func pathsFromGlobPattern(evaluationGlobPattern string) (evaluationFilePaths []string, err error) {
+	if filepath.Base(evaluationGlobPattern) != "evaluation.csv" {
+		return nil, pkgerrors.WithStack(pkgerrors.Errorf(`the path needs to end with "evaluation.csv" file, but found %q`, evaluationGlobPattern))
+	}
+
+	evaluationGlobFilePaths, err := filepath.Glob(evaluationGlobPattern)
+	if err != nil {
+		return nil, pkgerrors.WithStack(err)
+	} else if len(evaluationGlobFilePaths) == 0 {
+		return nil, pkgerrors.Errorf("no files matched the pattern %q", evaluationGlobPattern)
+	}
+
+	evaluationCSVFilePaths := map[string]bool{}
+	for _, evaluationCSVFilePath := range evaluationGlobFilePaths {
+		evaluationCSVFilePaths[evaluationCSVFilePath] = true
+	}
+
+	return maps.Keys(evaluationCSVFilePaths), nil
+}