class SpreadsheetParser # Based on file's extension opens file (used for importing) def self.open_spreadsheet(file) filename = file.original_filename file_path = file.path if file.class == Paperclip::Attachment && file.is_stored_on_s3? fa = file.fetch file_path = fa.path end case File.extname(filename) when '.csv' Roo::CSV.new(file_path, extension: :csv) when '.tsv' Roo::CSV.new(file_path, csv_options: { col_sep: "\t" }) when '.txt' # This assumption is based purely on biologist's habits Roo::CSV.new(file_path, csv_options: { col_sep: "\t" }) when '.xlsx' # Roo Excel parcel was replaced with Creek, but it can be enabled back, # just swap lines below. But only one can be enabled at the same time. # Roo::Excelx.new(file_path) Creek::Book.new(file_path).sheets[0] else raise TypeError end end def self.spreadsheet_enumerator(sheet) if sheet.is_a?(Roo::CSV) sheet elsif sheet.is_a?(Roo::Excelx) sheet.each_row_streaming else sheet.rows end end def self.first_two_rows(sheet) rows = spreadsheet_enumerator(sheet) header = [] columns = [] i = 1 rows.each do |row| # Creek XLSX parser returns Hash of the row, Roo - Array row = row.is_a?(Hash) ? row.values.map(&:to_s) : row.map(&:to_s) header = row if i == 1 && row columns = row if i == 2 && row i += 1 break if i > 2 end return header, columns end end