tiny-fast-gem 0.0.1

data/puma-8.0.2/lib/puma/cluster_accept_loop_delay.rb

@@ -0,0 +1,91 @@
+# frozen_string_literal: true
+
+module Puma
+  # Calculate a delay value for sleeping when running in clustered mode
+  #
+  # The main reason this is a class is so it can be unit tested independently.
+  # This makes modification easier in the future if we can encode properties of the
+  # delay into a test instead of relying on end-to-end testing only.
+  #
+  # This is an imprecise mechanism to address specific goals:
+  #
+  # - Evenly distribute requests across all workers at start
+  # - Evenly distribute CPU resources across all workers
+  #
+  # ## Goal: Distribute requests across workers at start
+  #
+  # There was a perf bug in Puma where one worker would wake up slightly before the rest and accept
+  # all the requests on the socket even though it didn't have enough resources to process all of them.
+  # This was originally fixed by never calling accept when a worker had more requests than threads
+  # already https://github.com/puma/puma/pull/3678/files/2736ebddb3fc8528e5150b5913fba251c37a8bf7#diff-a95f46e7ce116caddc9b9a9aa81004246d5210d5da5f4df90a818c780630166bL251-L291
+  #
+  # With the introduction of true keepalive support, there are two ways a request can come in:
+  # - A new request from a new client comes into the socket and it must be "accept"-ed
+  # - A keepalive request is served and the connection is retained. Another request is then accepted
+  #
+  # Ideally the server handles requests in the order they come in, and ideally it doesn't accept more requests than it can handle.
+  # These goals are contradictory, because when the server is at maximum capacity due to keepalive connections, it could mean we
+  # block all new requests, even if those came in before the new request on the older keepalive connection.
+  #
+  # ## Goal: Distribute CPU resources across all workers
+  #
+  # - This issue was opened https://github.com/puma/puma/issues/2078
+  #
+  # There are several entangled issues and it's not exactly clear what the root cause is, but the observable outcome
+  # was that performance was better with a small sleep, and that eventually became the default.
+  #
+  # An attempt to describe why this works is here: https://github.com/puma/puma/issues/2078#issuecomment-3287032470.
+  #
+  # Summarizing: The delay is for tuning the rate at which "accept" is called on the socket.
+  # Puma works by calling "accept" nonblock on the socket in a loop. When there are multiple workers
+  # (processes), they will "race" to accept a request at roughly the same rate. However, if one
+  # worker has all threads busy processing requests, then accepting a new request might "steal" it from
+  # a less busy worker. If a worker has no work to do, it should loop as fast as possible.
+  #
+  # ## Solution: Distribute requests across workers at start
+  #
+  # For now, both goals are framed as "load balancing" across workers (processes) and achieved through
+  # the same mechanism of sleeping longer to delay busier workers. Rather than the prior Puma 6.x
+  # and earlier behavior of using a binary on/off sleep value, we increase it an amount proportional
+  # to the load the server is under, capping the maximum delay to the scenario where all threads are busy
+  # and the todo list has reached a multiplier of the maximum number of threads.
+  #
+  # Private: API may change unexpectedly
+  class ClusterAcceptLoopDelay
+    attr_reader :max_delay
+
+    # Initialize happens once, `call` happens often. Perform global calculations here.
+    def initialize(
+      # Number of workers in the cluster
+      workers: ,
+      # Maximum delay in seconds i.e. 0.005 is 5 milliseconds
+      max_delay:
+    )
+      @on = max_delay > 0 && workers >= 2
+      @max_delay = max_delay.to_f
+
+      # Reach maximum delay when `max_threads * overload_multiplier` is reached in the system
+      @overload_multiplier = 25.0
+    end
+
+    def on?
+      @on
+    end
+
+    # We want the extreme values of this delay to be known (minimum and maximum) as well as
+    # a predictable curve between the two. i.e. no step functions or hard cliffs.
+    #
+    # Return value is always numeric. Returns 0 if there should be no delay.
+    def calculate(
+      # Number of threads working right now, plus number of requests in the todo list
+      busy_threads_plus_todo:,
+      # Maximum number of threads in the pool, note that the busy threads (alone) may go over this value at times
+      # if the pool needs to be reaped. The busy thread plus todo count may go over this value by a large amount.
+      max_threads:
+    )
+      max_value = @overload_multiplier * max_threads
+      # Approaches max delay when `busy_threads_plus_todo` approaches `max_value`
+      return max_delay * busy_threads_plus_todo.clamp(0, max_value) / max_value
+    end
+  end
+end

data/puma-8.0.2/lib/puma/commonlogger.rb

@@ -0,0 +1,115 @@
+# frozen_string_literal: true
+
+module Puma
+  # Rack::CommonLogger forwards every request to the given +app+, and
+  # logs a line in the
+  # {Apache common log format}[https://httpd.apache.org/docs/2.4/logs.html#common]
+  # to the +logger+.
+  #
+  # If +logger+ is nil, CommonLogger will fall back +rack.errors+, which is
+  # an instance of Rack::NullLogger.
+  #
+  # +logger+ can be any class, including the standard library Logger, and is
+  # expected to have either +write+ or +<<+ method, which accepts the CommonLogger::FORMAT.
+  # According to the SPEC, the error stream must also respond to +puts+
+  # (which takes a single argument that responds to +to_s+), and +flush+
+  # (which is called without arguments in order to make the error appear for
+  # sure)
+  class CommonLogger
+    # Common Log Format: https://httpd.apache.org/docs/2.4/logs.html#common
+    #
+    #   lilith.local - - [07/Aug/2006 23:58:02 -0400] "GET / HTTP/1.1" 500 -
+    #
+    #   %{%s - %s [%s] "%s %s%s %s" %d %s\n} %
+    FORMAT = %{%s - %s [%s] "%s %s%s %s" %d %s %0.4f\n}
+
+    HIJACK_FORMAT = %{%s - %s [%s] "%s %s%s %s" HIJACKED -1 %0.4f\n}
+
+    LOG_TIME_FORMAT = '%d/%b/%Y:%H:%M:%S %z'
+
+    CONTENT_LENGTH       = 'Content-Length' # should be lower case from app,
+                                            # Util::HeaderHash allows mixed
+    HTTP_X_FORWARDED_FOR = Const::HTTP_X_FORWARDED_FOR
+    PATH_INFO            = Const::PATH_INFO
+    QUERY_STRING         = Const::QUERY_STRING
+    REMOTE_ADDR          = Const::REMOTE_ADDR
+    REMOTE_USER          = 'REMOTE_USER'
+    REQUEST_METHOD       = Const::REQUEST_METHOD
+    SERVER_PROTOCOL      = Const::SERVER_PROTOCOL
+
+    def initialize(app, logger=nil)
+      @app = app
+      @logger = logger
+    end
+
+    def call(env)
+      began_at = Time.now
+      status, header, body = @app.call(env)
+      header = Util::HeaderHash.new(header)
+
+      # If we've been hijacked, then output a special line
+      if env['rack.hijack_io']
+        log_hijacking(env, 'HIJACK', header, began_at)
+      else
+        ary = env['rack.after_reply']
+        ary << lambda { log(env, status, header, began_at) }
+      end
+
+      [status, header, body]
+    end
+
+    private
+
+    def log_hijacking(env, status, header, began_at)
+      now = Time.now
+
+      msg = HIJACK_FORMAT % [
+        env[HTTP_X_FORWARDED_FOR] || env[REMOTE_ADDR] || "-",
+        env[REMOTE_USER] || "-",
+        now.strftime(LOG_TIME_FORMAT),
+        env[REQUEST_METHOD],
+        env[PATH_INFO],
+        env[QUERY_STRING].empty? ? "" : "?#{env[QUERY_STRING]}",
+        env[SERVER_PROTOCOL],
+        now - began_at ]
+
+      write(msg)
+    end
+
+    def log(env, status, header, began_at)
+      now = Time.now
+      length = extract_content_length(header)
+
+      msg = FORMAT % [
+        env[HTTP_X_FORWARDED_FOR] || env[REMOTE_ADDR] || "-",
+        env[REMOTE_USER] || "-",
+        now.strftime(LOG_TIME_FORMAT),
+        env[REQUEST_METHOD],
+        env[PATH_INFO],
+        env[QUERY_STRING].empty? ? "" : "?#{env[QUERY_STRING]}",
+        env[SERVER_PROTOCOL],
+        status.to_s[0..3],
+        length,
+        now - began_at ]
+
+      write(msg)
+    end
+
+    def write(msg)
+      logger = @logger || env['rack.errors']
+
+      # Standard library logger doesn't support write but it supports << which actually
+      # calls to write on the log device without formatting
+      if logger.respond_to?(:write)
+        logger.write(msg)
+      else
+        logger << msg
+      end
+    end
+
+    def extract_content_length(headers)
+      value = headers[CONTENT_LENGTH] or return '-'
+      value.to_s == '0' ? '-' : value
+    end
+  end
+end