require 'socket' require 'net/https' require 'net/http' require 'logger' require 'zlib' require 'stringio' # The NewRelic Agent collects performance data from ruby applications # in realtime as the application runs, and periodically sends that # data to the NewRelic server. module NewRelic module Agent # The Agent is a singleton that is instantiated when the plugin is # activated. class Agent # Specifies the version of the agent's communication protocol with # the NewRelic hosted site. PROTOCOL_VERSION = 8 attr_reader :obfuscator attr_reader :stats_engine attr_reader :transaction_sampler attr_reader :error_collector attr_reader :record_sql attr_reader :histogram attr_reader :metric_ids # Should only be called by NewRelic::Control def self.instance @instance ||= self.new end # This method is deprecated. Use NewRelic::Agent.manual_start def manual_start(ignored=nil, also_ignored=nil) raise "This method no longer supported. Instead use the class method NewRelic::Agent.manual_start" end # This method should be called in a forked process after a fork. # It assumes the parent process started the agent. It clears any # metrics carried over from the parent process, and restarts # the worker thread and sampler threads. # # In addition, if the parent process had not connected to the server # because it was a Passenger or Unicorn spawner, it will # initiate the connection and create a new agent run. # def after_fork # @connected gets false after we fail to connect or have an error # connecting. @connected has nil if we haven't finished trying to connect. # or we didn't attempt a connection because this is the master process log.debug "Agent received after_fork notice in #$$: [#{control.agent_enabled?}; monitor=#{control.monitor_mode?}; connected: #{@connected.inspect}; thread=#{@worker_thread.inspect}]" return if !control.agent_enabled? or !control.monitor_mode? or @connected == false or @worker_thread && @worker_thread.alive? log.debug "Detected that the worker thread is not running in #$$. Restarting." # Clear out stats that are left over from parent process reset_stats start_worker_thread @stats_engine.start_sampler_thread end # True if we have initialized and completed 'start' def started? @started end # Attempt a graceful shutdown of the agent. def shutdown return if not started? if @worker_loop @worker_loop.stop log.debug "Starting Agent shutdown" # if litespeed, then ignore all future SIGUSR1 - it's # litespeed trying to shut us down if control.dispatcher == :litespeed Signal.trap("SIGUSR1", "IGNORE") Signal.trap("SIGTERM", "IGNORE") end begin graceful_disconnect rescue => e log.error e log.error e.backtrace.join("\n") end end @started = nil end def start_transaction @stats_engine.start_transaction end def end_transaction @stats_engine.end_transaction end def set_record_sql(should_record) prev = Thread::current[:record_sql] Thread::current[:record_sql] = should_record prev.nil? || prev end def set_record_tt(should_record) prev = Thread::current[:record_tt] Thread::current[:record_tt] = should_record prev.nil? || prev end # Push flag indicating whether we should be tracing in this # thread. def push_trace_execution_flag(should_trace=false) (Thread.current[:newrelic_untraced] ||= []) << should_trace end # Pop the current trace execution status. Restore trace execution status # to what it was before we pushed the current flag. def pop_trace_execution_flag Thread.current[:newrelic_untraced].pop if Thread.current[:newrelic_untraced] end def set_sql_obfuscator(type, &block) if type == :before @obfuscator = NewRelic::ChainedCall.new(block, @obfuscator) elsif type == :after @obfuscator = NewRelic::ChainedCall.new(@obfuscator, block) elsif type == :replace @obfuscator = block else fail "unknown sql_obfuscator type #{type}" end end def log NewRelic::Agent.logger end # Start up the agent. This verifies that the agent_enabled? is # true and initializes the sampler based on the current # controluration settings. Then it will fire up the background # thread for sending data to the server if applicable. def start if started? control.log! "Agent Started Already!", :error return end return if !control.agent_enabled? @started = true @local_host = determine_host if control.dispatcher.nil? || control.dispatcher.to_s.empty? log.info "No dispatcher detected." else log.info "Dispatcher: #{control.dispatcher.to_s}" end log.info "Application: #{control.app_names.join(", ")}" unless control.app_names.empty? sampler_config = control.fetch('transaction_tracer', {}) @should_send_samples = sampler_config.fetch('enabled', true) log.info "Transaction tracing not enabled." if not @should_send_samples @record_sql = sampler_config.fetch('record_sql', :obfuscated).to_sym # use transaction_threshold: 4.0 to force the TT collection # threshold to 4 seconds # use transaction_threshold: apdex_f to use your apdex t value # multiplied by 4 # undefined transaction_threshold defaults to 2.0 apdex_f = 4 * NewRelic::Control.instance.apdex_t @slowest_transaction_threshold = sampler_config.fetch('transaction_threshold', 2.0) if @slowest_transaction_threshold =~ /apdex_f/i @slowest_transaction_threshold = apdex_f end @slowest_transaction_threshold = @slowest_transaction_threshold.to_f @explain_threshold = sampler_config.fetch('explain_threshold', 0.5).to_f @explain_enabled = sampler_config.fetch('explain_enabled', true) @random_sample = sampler_config.fetch('random_sample', false) log.warn "Agent is configured to send raw SQL to RPM service" if @record_sql == :raw # Initialize transaction sampler @transaction_sampler.random_sampling = @random_sample if control.monitor_mode? if !control.license_key control.log! "No license key found. Please edit your newrelic.yml file and insert your license key.", :error elsif control.license_key.length != 40 control.log! "Invalid license key: #{control.license_key}", :error else # Do the connect in the foreground if we are in sync mode NewRelic::Agent.disable_all_tracing { connect(false) } if control.sync_startup # Start the event loop and initiate connection if necessary start_worker_thread # Our shutdown handler needs to run after other shutdown handlers # that may be doing things like running the app (hello sinatra). if RUBY_VERSION =~ /rubinius/i list = at_exit { shutdown } # move the shutdown handler to the front of the list, to execute last: list.unshift(list.pop) elsif !defined?(JRuby) or !defined(Sinatra::Application) at_exit { at_exit { shutdown } } end end end control.log! "New Relic RPM Agent #{NewRelic::VERSION::STRING} Initialized: pid = #$$" control.log! "Agent Log found in #{NewRelic::Control.instance.log_file}" if NewRelic::Control.instance.log_file end private def collector @collector ||= control.server end # Try to launch the worker thread and connect to the server def start_worker_thread @worker_loop = WorkerLoop.new log.debug "Creating RPM worker thread." @worker_thread = Thread.new do begin NewRelic::Agent.disable_all_tracing do # We try to connect. If this returns false that means # the server rejected us for a licensing reason and we should # just exit the thread. If it returns nil # that means it didn't try to connect because we're in the master connect if not @connected if @connected # disable transaction sampling if disabled by the server and we're not in dev mode if !control.developer_mode? && !@should_send_samples @transaction_sampler.disable end control.log! "Reporting performance data every #{@report_period} seconds." log.debug "Running worker loop" # note if the agent attempts to report more frequently than allowed by the server # the server will start dropping data. @worker_loop.run(@report_period) do harvest_and_send_timeslice_data harvest_and_send_slowest_sample if @should_send_samples harvest_and_send_errors if error_collector.enabled end @connected = true end end rescue NewRelic::Agent::ForceRestartException => e log.info e.message # disconnect and start over. # clear the stats engine reset_stats @metric_ids = {} @connected = nil # Wait a short time before trying to reconnect sleep 30 retry rescue NewRelic::Agent::ForceDisconnectException => e # when a disconnect is requested, stop the current thread, which # is the worker thread that gathers data and talks to the # server. log.error "RPM forced this agent to disconnect (#{e.message})" @connected = false rescue NewRelic::Agent::ServerConnectionException => e control.log! "Unable to establish connection with the server. Run with log level set to debug for more information." log.debug("#{e.class.name}: #{e.message}\n#{e.backtrace.first}") @connected = false rescue Exception => e log.error "Terminating worker loop: #{e.class.name}: #{e}\n #{e.backtrace.join("\n ")}" @connected = false end # begin end # thread new @worker_thread['newrelic_label'] = 'Worker Loop' end def control NewRelic::Control.instance end def initialize @launch_time = Time.now @metric_ids = {} @histogram = NewRelic::Histogram.new(NewRelic::Control.instance.apdex_t / 10) @stats_engine = NewRelic::Agent::StatsEngine.new @transaction_sampler = NewRelic::Agent::TransactionSampler.new @stats_engine.transaction_sampler = @transaction_sampler @error_collector = NewRelic::Agent::ErrorCollector.new @request_timeout = NewRelic::Control.instance.fetch('timeout', 2 * 60) @last_harvest_time = Time.now @obfuscator = method(:default_sql_obfuscator) end # Connect to the server and validate the license. If successful, # @connected has true when finished. If not successful, you can # keep calling this. Return false if we could not establish a # connection with the server and we should not retry, such as if # there's a bad license key. # # Set keep_retrying=false to disable retrying and return asap, such as when # invoked in the foreground. Otherwise this runs until a successful # connection is made, or the server rejects us. def connect(keep_retrying = true) # Don't reconnect if we've already got a connection. return if @connected # wait a few seconds for the web server to boot, necessary in development connect_retry_period = keep_retrying ? 10 : 0 connect_attempts = 0 @agent_id = nil begin sleep connect_retry_period.to_i # Running in the Passenger or Unicorn spawners? if $0 =~ /ApplicationSpawner|master/ log.debug "Process is master spawner (#$0) -- don't connect to RPM service" @connected = nil return else log.debug "Connecting Process to RPM: #$0" end environment = control['send_environment_info'] != false ? control.local_env.snapshot : [] log.debug "Connecting with validation seed/token: #{control.validate_seed}/#{control.validate_token}" if control.validate_seed @agent_id ||= invoke_remote :start, @local_host, { :pid => $$, :launch_time => @launch_time.to_f, :agent_version => NewRelic::VERSION::STRING, :environment => environment, :settings => control.settings, :validate_seed => control.validate_seed, :validate_token => control.validate_token } host = invoke_remote(:get_redirect_host) @collector = control.server_from_host(host) if host @report_period = invoke_remote :get_data_report_period, @agent_id control.log! "Connected to NewRelic Service at #{@collector}" log.debug "Agent ID = #{@agent_id}." # Ask the server for permission to send transaction samples. # determined by subscription license. @should_send_samples &&= invoke_remote :should_collect_samples, @agent_id if @should_send_samples sampling_rate = invoke_remote :sampling_rate, @agent_id if @random_sample @transaction_sampler.sampling_rate = sampling_rate log.info "Transaction sample rate: #{@transaction_sampler.sampling_rate}" if sampling_rate log.info "Transaction tracing threshold is #{@slowest_transaction_threshold} seconds." end # Ask for permission to collect error data error_collector.enabled &&= invoke_remote(:should_collect_errors, @agent_id) log.info "Transaction traces will be sent to the RPM service." if @should_send_samples log.info "Errors will be sent to the RPM service." if error_collector.enabled @connected_pid = $$ @connected = true rescue NewRelic::Agent::LicenseException => e control.log! e.message, :error control.log! "Visit NewRelic.com to obtain a valid license key, or to upgrade your account." @connected = false rescue Timeout::Error, StandardError => e log.info "Unable to establish connection with New Relic RPM Service at #{control.server}" unless e.instance_of? NewRelic::Agent::ServerConnectionException log.error e.message log.debug e.backtrace.join("\n") end # retry logic if keep_retrying connect_attempts += 1 case connect_attempts when 1..2 connect_retry_period, period_msg = 60, "1 minute" when 3..5 connect_retry_period, period_msg = 60 * 2, "2 minutes" else connect_retry_period, period_msg = 5 * 60, "5 minutes" end log.info "Will re-attempt in #{period_msg}" retry else @connected = nil end end end def determine_host Socket.gethostname end def determine_home_directory control.root end def reset_stats @stats_engine.reset_stats @unsent_errors = [] @traces = nil @unsent_timeslice_data = {} @last_harvest_time = Time.now @launch_time = Time.now @histogram = NewRelic::Histogram.new(NewRelic::Control.instance.apdex_t / 10) end def harvest_and_send_timeslice_data NewRelic::Agent::BusyCalculator.harvest_busy now = Time.now @unsent_timeslice_data ||= {} @unsent_timeslice_data = @stats_engine.harvest_timeslice_data(@unsent_timeslice_data, @metric_ids) begin # In this version of the protocol, we get back an assoc array of spec to id. metric_ids = invoke_remote(:metric_data, @agent_id, @last_harvest_time.to_f, now.to_f, @unsent_timeslice_data.values) rescue Timeout::Error # assume that the data was received. chances are that it was metric_ids = nil end metric_ids.each do | spec, id | @metric_ids[spec] = id end if metric_ids log.debug "#{now}: sent #{@unsent_timeslice_data.length} timeslices (#{@agent_id}) in #{Time.now - now} seconds" # if we successfully invoked this web service, then clear the unsent message cache. @unsent_timeslice_data = {} @last_harvest_time = now # handle_messages # note - exceptions are logged in invoke_remote. If an exception is encountered here, # then the metric data is downsampled for another timeslices end def harvest_and_send_slowest_sample @traces = @transaction_sampler.harvest(@traces, @slowest_transaction_threshold) unless @traces.empty? now = Time.now log.debug "Sending (#{@traces.length}) transaction traces" begin # take the traces and prepare them for sending across the # wire. This includes gathering SQL explanations, stripping # out stack traces, and normalizing SQL. note that we # explain only the sql statements whose segments' execution # times exceed our threshold (to avoid unnecessary overhead # of running explains on fast queries.) traces = @traces.collect {|trace| trace.prepare_to_send(:explain_sql => @explain_threshold, :record_sql => @record_sql, :keep_backtraces => true, :explain_enabled => @explain_enabled)} invoke_remote :transaction_sample_data, @agent_id, traces rescue PostTooBigException # we tried to send too much data, drop the first trace and # try again retry if @traces.shift end log.debug "Sent slowest sample (#{@agent_id}) in #{Time.now - now} seconds" end # if we succeed sending this sample, then we don't need to keep # the slowest sample around - it has been sent already and we # can collect the next one @traces = nil # note - exceptions are logged in invoke_remote. If an # exception is encountered here, then the slowest sample of is # determined of the entire period since the last reported # sample. end def harvest_and_send_errors @unsent_errors = @error_collector.harvest_errors(@unsent_errors) if @unsent_errors && @unsent_errors.length > 0 log.debug "Sending #{@unsent_errors.length} errors" begin invoke_remote :error_data, @agent_id, @unsent_errors rescue PostTooBigException @unsent_errors.shift retry end # if the remote invocation fails, then we never clear # @unsent_errors, and therefore we can re-attempt to send on # the next heartbeat. Note the error collector maxes out at # 20 instances to prevent leakage @unsent_errors = [] end end def compress_data(object) dump = Marshal.dump(object) # this checks to make sure mongrel won't choke on big uploads check_post_size(dump) # we currently optimize for CPU here since we get roughly a 10x # reduction in message size with this, and CPU overhead is at a # premium. For extra-large posts, we use the higher compression # since otherwise it actually errors out. dump_size = dump.size # small payloads don't need compression return [dump, 'identity'] if dump_size < 2000 # medium payloads get fast compression, to save CPU # big payloads get all the compression possible, to stay under # the 2,000,000 byte post threshold compression = dump_size < 2000000 ? Zlib::BEST_SPEED : Zlib::BEST_COMPRESSION [Zlib::Deflate.deflate(dump, compression), 'deflate'] end def check_post_size(post_string) # TODO: define this as a config option on the server side return if post_string.size < control.post_size_limit log.warn "Tried to send too much data: #{post_string.size} bytes" raise PostTooBigException end def send_request(opts) request = Net::HTTP::Post.new(opts[:uri], 'CONTENT-ENCODING' => opts[:encoding], 'ACCEPT-ENCODING' => 'gzip', 'HOST' => opts[:collector].name) request.content_type = "application/octet-stream" request.body = opts[:data] log.debug "Connect to #{opts[:collector]}#{opts[:uri]}" response = nil http = control.http_connection(collector) begin timeout(@request_timeout) do response = http.request(request) end rescue Timeout::Error log.warn "Timed out trying to post data to RPM (timeout = #{@request_timeout} seconds)" unless @request_timeout < 30 raise end if response.is_a? Net::HTTPServiceUnavailable raise NewRelic::Agent::ServerConnectionException, "Service unavailable: #{response.body || response.message}" elsif response.is_a? Net::HTTPGatewayTimeOut log.debug("Timed out getting response: #{response.message}") raise Timeout::Error, response.message elsif !(response.is_a? Net::HTTPSuccess) raise NewRelic::Agent::ServerConnectionException, "Unexpected response from server: #{response.code}: #{response.message}" end response end def decompress_response(response) if response['content-encoding'] != 'gzip' log.debug "Uncompressed content returned" return response.body end log.debug "Decompressing return value" i = Zlib::GzipReader.new(StringIO.new(response.body)) i.read end def check_for_exception(response) dump = decompress_response(response) value = Marshal.load(dump) raise value if value.is_a? Exception value end def remote_method_uri(method) uri = "/agent_listener/#{PROTOCOL_VERSION}/#{control.license_key}/#{method}" uri << "?run_id=#{@agent_id}" if @agent_id uri end # send a message via post def invoke_remote(method, *args) #determines whether to zip the data or send plain post_data, encoding = compress_data(args) response = send_request({:uri => remote_method_uri(method), :encoding => encoding, :collector => collector, :data => post_data}) # raises the right exception if the remote server tells it to die return check_for_exception(response) rescue NewRelic::Agent::ForceRestartException => e log.info e.message raise rescue SystemCallError, SocketError => e # These include Errno connection errors raise NewRelic::Agent::ServerConnectionException, "Recoverable error connecting to the server: #{e}" end def graceful_disconnect if @connected begin log.debug "Sending graceful shutdown message to #{control.server}" @request_timeout = 10 log.debug "Flushing unsent metric data to server" @worker_loop.run_task if @connected_pid == $$ log.debug "Sending RPM service agent run shutdown message" invoke_remote :shutdown, @agent_id, Time.now.to_f else log.debug "This agent connected from #{@connected_pid}--not sending shutdown" end log.debug "Graceful disconnect complete" rescue Timeout::Error, StandardError end else log.debug "Bypassing graceful disconnect - agent not connected" end end def default_sql_obfuscator(sql) sql = sql.dup # This is hardly readable. Use the unit tests. # remove single quoted strings: sql.gsub!(/'(.*?[^\\'])??'(?!')/, '?') # remove double quoted strings: sql.gsub!(/"(.*?[^\\"])??"(?!")/, '?') # replace all number literals sql.gsub!(/\d+/, "?") sql end end end end