package yqlib import ( "bufio" "fmt" "testing" "github.com/mikefarah/yq/v4/test" ) const csvSimple = `name,numberOfCats,likesApples,height Gary,1,true,168.8 Samantha's Rabbit,2,false,-188.8 ` const expectedUpdatedSimpleCsv = `name,numberOfCats,likesApples,height Gary,3,true,168.8 Samantha's Rabbit,2,false,-188.8 ` const csvSimpleShort = `Name,Number of Cats Gary,1 Samantha's Rabbit,2 ` const tsvSimple = `name numberOfCats likesApples height Gary 1 true 168.8 Samantha's Rabbit 2 false -188.8 ` const expectedYamlFromCSV = `- name: Gary numberOfCats: 1 likesApples: true height: 168.8 - name: Samantha's Rabbit numberOfCats: 2 likesApples: false height: -188.8 ` const expectedYamlFromCSVMissingData = `- name: Gary numberOfCats: 1 height: 168.8 - name: Samantha's Rabbit height: -188.8 likesApples: false ` const csvSimpleMissingData = `name,numberOfCats,height Gary,1,168.8 Samantha's Rabbit,,-188.8 ` const csvTestSimpleYaml = `- [i, like, csv] - [because, excel, is, cool]` const expectedSimpleCsv = `i,like,csv because,excel,is,cool ` const tsvTestExpectedSimpleCsv = `i like csv because excel is cool ` var csvScenarios = []formatScenario{ { description: "Encode CSV simple", input: csvTestSimpleYaml, expected: expectedSimpleCsv, scenarioType: "encode-csv", }, { description: "Encode TSV simple", input: csvTestSimpleYaml, expected: tsvTestExpectedSimpleCsv, scenarioType: "encode-tsv", }, { description: "Encode Empty", skipDoc: true, input: `[]`, expected: "", scenarioType: "encode-csv", }, { description: "Comma in value", skipDoc: true, input: `["comma, in, value", things]`, expected: "\"comma, in, value\",things\n", scenarioType: "encode-csv", }, { description: "Encode array of objects to csv", input: expectedYamlFromCSV, expected: csvSimple, scenarioType: "encode-csv", }, { description: "Encode array of objects to custom csv format", subdescription: "Add the header row manually, then the we convert each object into an array of values - resulting in an array of arrays. Pick the columns and call the header whatever you like.", input: expectedYamlFromCSV, expected: csvSimpleShort, expression: `[["Name", "Number of Cats"]] + [.[] | [.name, .numberOfCats ]]`, scenarioType: "encode-csv", }, { description: "Encode array of objects to csv - missing fields behaviour", subdescription: "First entry is used to determine the headers, and it is missing 'likesApples', so it is not included in the csv. Second entry does not have 'numberOfCats' so that is blank", input: expectedYamlFromCSVMissingData, expected: csvSimpleMissingData, scenarioType: "encode-csv", }, { description: "Parse CSV into an array of objects", subdescription: "First row is assumed to be the header row.", input: csvSimple, expected: expectedYamlFromCSV, scenarioType: "decode-csv-object", }, { description: "Parse TSV into an array of objects", subdescription: "First row is assumed to be the header row.", input: tsvSimple, expected: expectedYamlFromCSV, scenarioType: "decode-tsv-object", }, { description: "Round trip", input: csvSimple, expected: expectedUpdatedSimpleCsv, expression: `(.[] | select(.name == "Gary") | .numberOfCats) = 3`, scenarioType: "roundtrip-csv", }, } func testCSVScenario(t *testing.T, s formatScenario) { switch s.scenarioType { case "encode-csv": test.AssertResultWithContext(t, s.expected, processFormatScenario(s, NewYamlDecoder(), NewCsvEncoder(',')), s.description) case "encode-tsv": test.AssertResultWithContext(t, s.expected, processFormatScenario(s, NewYamlDecoder(), NewCsvEncoder('\t')), s.description) case "decode-csv-object": test.AssertResultWithContext(t, s.expected, processFormatScenario(s, NewCSVObjectDecoder(','), NewYamlEncoder(2, false, true, true)), s.description) case "decode-tsv-object": test.AssertResultWithContext(t, s.expected, processFormatScenario(s, NewCSVObjectDecoder('\t'), NewYamlEncoder(2, false, true, true)), s.description) case "roundtrip-csv": test.AssertResultWithContext(t, s.expected, processFormatScenario(s, NewCSVObjectDecoder(','), NewCsvEncoder(',')), s.description) default: panic(fmt.Sprintf("unhandled scenario type %q", s.scenarioType)) } } func documentCSVDecodeObjectScenario(t *testing.T, w *bufio.Writer, s formatScenario, formatType string) { writeOrPanic(w, fmt.Sprintf("## %v\n", s.description)) if s.subdescription != "" { writeOrPanic(w, s.subdescription) writeOrPanic(w, "\n\n") } writeOrPanic(w, fmt.Sprintf("Given a sample.%v file of:\n", formatType)) writeOrPanic(w, fmt.Sprintf("```%v\n%v\n```\n", formatType, s.input)) writeOrPanic(w, "then\n") writeOrPanic(w, fmt.Sprintf("```bash\nyq -p=%v sample.%v\n```\n", formatType, formatType)) writeOrPanic(w, "will output\n") separator := ',' if formatType == "tsv" { separator = '\t' } writeOrPanic(w, fmt.Sprintf("```yaml\n%v```\n\n", processFormatScenario(s, NewCSVObjectDecoder(separator), NewYamlEncoder(s.indent, false, true, true))), ) } func documentCSVEncodeScenario(w *bufio.Writer, s formatScenario, formatType string) { writeOrPanic(w, fmt.Sprintf("## %v\n", s.description)) if s.subdescription != "" { writeOrPanic(w, s.subdescription) writeOrPanic(w, "\n\n") } writeOrPanic(w, "Given a sample.yml file of:\n") writeOrPanic(w, fmt.Sprintf("```yaml\n%v\n```\n", s.input)) writeOrPanic(w, "then\n") expression := s.expression if expression != "" { writeOrPanic(w, fmt.Sprintf("```bash\nyq -o=%v '%v' sample.yml\n```\n", formatType, expression)) } else { writeOrPanic(w, fmt.Sprintf("```bash\nyq -o=%v sample.yml\n```\n", formatType)) } writeOrPanic(w, "will output\n") separator := ',' if formatType == "tsv" { separator = '\t' } writeOrPanic(w, fmt.Sprintf("```%v\n%v```\n\n", formatType, processFormatScenario(s, NewYamlDecoder(), NewCsvEncoder(separator))), ) } func documentCSVRoundTripScenario(w *bufio.Writer, s formatScenario, formatType string) { writeOrPanic(w, fmt.Sprintf("## %v\n", s.description)) if s.subdescription != "" { writeOrPanic(w, s.subdescription) writeOrPanic(w, "\n\n") } writeOrPanic(w, fmt.Sprintf("Given a sample.%v file of:\n", formatType)) writeOrPanic(w, fmt.Sprintf("```%v\n%v\n```\n", formatType, s.input)) writeOrPanic(w, "then\n") expression := s.expression if expression != "" { writeOrPanic(w, fmt.Sprintf("```bash\nyq -p=%v -o=%v '%v' sample.%v\n```\n", formatType, formatType, expression, formatType)) } else { writeOrPanic(w, fmt.Sprintf("```bash\nyq -p=%v -o=%v sample.%v\n```\n", formatType, formatType, formatType)) } writeOrPanic(w, "will output\n") separator := ',' if formatType == "tsv" { separator = '\t' } writeOrPanic(w, fmt.Sprintf("```%v\n%v```\n\n", formatType, processFormatScenario(s, NewCSVObjectDecoder(separator), NewCsvEncoder(separator))), ) } func documentCSVScenario(t *testing.T, w *bufio.Writer, i interface{}) { s := i.(formatScenario) if s.skipDoc { return } switch s.scenarioType { case "encode-csv": documentCSVEncodeScenario(w, s, "csv") case "encode-tsv": documentCSVEncodeScenario(w, s, "tsv") case "decode-csv-object": documentCSVDecodeObjectScenario(t, w, s, "csv") case "decode-tsv-object": documentCSVDecodeObjectScenario(t, w, s, "tsv") case "roundtrip-csv": documentCSVRoundTripScenario(w, s, "csv") default: panic(fmt.Sprintf("unhandled scenario type %q", s.scenarioType)) } } func TestCSVScenarios(t *testing.T) { for _, tt := range csvScenarios { testCSVScenario(t, tt) } genericScenarios := make([]interface{}, len(csvScenarios)) for i, s := range csvScenarios { genericScenarios[i] = s } documentScenarios(t, "usage", "csv-tsv", genericScenarios, documentCSVScenario) }