diff --git a/lib/logtail/log_devices/http.rb b/lib/logtail/log_devices/http.rb
index 54c2048..926ab1e 100755
--- a/lib/logtail/log_devices/http.rb
+++ b/lib/logtail/log_devices/http.rb
@@ -85,6 +85,9 @@ def initialize(source_token, options = {})
         @flush_continuously = options[:flush_continuously] != false
         @flush_interval = options[:flush_interval] || 2 # 2 seconds
         @requests_per_conn = options[:requests_per_conn] || 2_500
+        # The process that owns the queues and threads, see {#reset_if_forked}
+        @pid = Process.pid
+        @fork_lock = Mutex.new
         @msg_queue = FlushableDroppingSizedQueue.new(@batch_size)
         @request_queue = options[:request_queue] || FlushableDroppingSizedQueue.new(25)
         @successive_error_count = 0
@@ -107,6 +110,7 @@ def write(msg)
         # Strings, e.g. from a plain ::Logger writing to this device, are sent as info lines.
         msg = LogEntry.new(:info, Time.now, nil, msg.to_s.chomp, nil, nil) unless msg.is_a?(LogEntry)
         return unless Logtail.config.send_to_better_stack?(msg)
+        reset_if_forked
 
         @msg_queue.enq(msg)
         # No thread delivers what is written after #close, e.g. by an at_exit hook that runs
@@ -133,17 +137,25 @@ def write(msg)
       end
 
       # Flush all log messages in the buffer synchronously. This method will not return
-      # until delivery of the messages has been successful. If you want to flush
+      # until delivery of the messages has been successful, or about 5 seconds have passed.
+      # When no outlet thread runs (`flush_continuously: false`, or a forked child that hasn't
+      # logged yet), the messages are delivered in the calling thread. If you want to flush
       # asynchronously see {#flush_async}.
       def flush
+        reset_if_forked
         flush_async
-        wait_on_request_queue
+        if @request_outlet_thread && @request_outlet_thread.alive?
+          wait_on_request_queue
+        else
+          deliver_synchronously(dequeue_requests)
+        end
         true
       end
 
       # Closes the log device, cleans up, and attempts one last delivery. Closing it again does
       # nothing; lines written after it are delivered right away (see {#write}).
       def close
+        reset_if_forked
         return if @closed
         @closed = true
 
@@ -226,6 +238,38 @@ def ensure_flush_threads_are_started
           end
         end
 
+        # The queues and threads belong to the process that created them. After a fork, the
+        # parent still delivers the lines it buffered, so a child that kept them would send them
+        # again, and the parent's threads don't run in the child. The child starts over with
+        # empty queues and starts its own threads once it logs, also when the parent closed the
+        # device before forking.
+        def reset_if_forked
+          return if @pid == Process.pid
+
+          @fork_lock.synchronize do
+            return if @pid == Process.pid
+
+            @msg_queue = FlushableDroppingSizedQueue.new(@batch_size)
+            # The request queue can be a SizedQueue passed as the :request_queue option
+            @request_queue.respond_to?(:flush) ? @request_queue.flush : @request_queue.clear
+            @flush_thread = @request_outlet_thread = nil
+            @requests_in_flight = 0
+            @reconnect_wait = INITIAL_RECONNECT_WAIT
+            @closed = @late_delivery_failed = false
+            @pid = Process.pid
+          end
+        end
+
+        # Takes the queued requests off the request queue, for {#flush} when no outlet thread
+        # runs. It checks the size first because a SizedQueue (see :request_queue) blocks when empty.
+        def dequeue_requests
+          requests = []
+          while @request_queue.size > 0 && (request_attempt = @request_queue.deq)
+            requests << request_attempt.request
+          end
+          requests
+        end
+
         # Builds an HTTP request based on the current messages queued.
         def build_request(msgs)
           path = '/'
@@ -447,8 +491,8 @@ def deliver_synchronously(requests)
         # Waits on the request queue. This is used in {#flush} to ensure
         # the log data has been delivered before returning.
         def wait_on_request_queue
-          # Wait 20 seconds
-          40.times do |i|
+          # Wait 5 seconds
+          10.times do |i|
             if @request_queue.size == 0 && @requests_in_flight == 0
               Logtail::Config.instance.debug { "Request queue is empty and no requests are in flight, finish waiting" }
               return true
diff --git a/lib/logtail/logger.rb b/lib/logtail/logger.rb
index 391ea1a..db1fa87 100755
--- a/lib/logtail/logger.rb
+++ b/lib/logtail/logger.rb
@@ -220,6 +220,21 @@ def level=(value)
       super
     end
 
+    # Delivers what was logged so far before it returns, waiting about 5 seconds at most with
+    # the HTTP log device. Call it before a process ends with `exit!`, which skips the at_exit
+    # hook that delivers the rest, as Resque's forked job processes do.
+    #
+    # Rails calls `flush` on Rails.logger after every request (ActiveSupport::LogSubscriber.flush_all!).
+    # Waiting there would hold up every request, so that call returns right away and the lines
+    # are delivered in the background as usual.
+    def flush
+      return true if caller_locations(1, 10).any? { |location| location.base_label == "flush_all!" }
+
+      @logdev.dev.flush if @logdev && @logdev.dev.respond_to?(:flush)
+      @extra_loggers.each { |logger| logger.flush if logger.respond_to?(:flush) }
+      true
+    end
+
     # @private
     def with_context(context, &block)
       Logtail::CurrentContext.with(context, &block)
diff --git a/spec/logtail/log_devices/http_fork_spec.rb b/spec/logtail/log_devices/http_fork_spec.rb
new file mode 100644
index 0000000..da47dee
--- /dev/null
+++ b/spec/logtail/log_devices/http_fork_spec.rb
@@ -0,0 +1,85 @@
+require "spec_helper"
+
+# Forking servers and job runners (Puma, Unicorn, Resque) fork after the app logged. These run
+# logtail in a real process that forks, and count what reaches a local ingesting server.
+describe Logtail::LogDevices::HTTP, "after a fork" do
+  let(:ingest) { LocalIngestServer.new }
+
+  before do
+    skip "needs fork" if !Process.respond_to?(:fork) || RUBY_ENGINE == "truffleruby"
+  end
+
+  after { ingest.stop }
+
+  it "delivers the lines the parent logged before forking once, not once per process" do
+    result = run_ruby(<<-RUBY)
+      require "logtail"
+      logger = Logtail::Logger.new(Logtail::LogDevices::HTTP.new("token", flush_interval: 60, #{ingest.device_options}))
+      logger.info("parent line before the fork")
+      2.times.map { |n| fork { logger.info("child line " + n.to_s) } }.each { |pid| Process.wait(pid) }
+    RUBY
+
+    expect(result.status).to be_success, result.stderr
+    expect(ingest.messages).to contain_exactly("parent line before the fork", "child line 0", "child line 1")
+  end
+
+  it "delivers the lines a forked child logs" do
+    result = run_ruby(<<-RUBY)
+      require "logtail"
+      logger = Logtail::Logger.new(Logtail::LogDevices::HTTP.new("token", flush_interval: 60, #{ingest.device_options}))
+      Process.wait(fork { 3.times { |n| logger.info("child line " + n.to_s) } })
+      logger.info("parent line after the fork")
+    RUBY
+
+    expect(result.status).to be_success, result.stderr
+    expect(ingest.messages).to contain_exactly("child line 0", "child line 1", "child line 2", "parent line after the fork")
+  end
+
+  it "lets a child that hasn't logged exit right away, without waiting on the parent's lines" do
+    result = run_ruby(<<-RUBY)
+      require "logtail"
+      logger = Logtail::Logger.new(Logtail::LogDevices::HTTP.new("token", flush_interval: 60, #{ingest.device_options}))
+      logger.info("parent line before the fork")
+      forking = Process.clock_gettime(Process::CLOCK_MONOTONIC)
+      Process.wait(fork {})
+      puts Process.clock_gettime(Process::CLOCK_MONOTONIC) - forking
+    RUBY
+
+    expect(result.stdout.to_f).to be < 5
+    expect(ingest.messages).to eq(["parent line before the fork"])
+  end
+
+  it "batches a child's lines as usual when the parent closed the device before forking" do
+    result = run_ruby(<<-RUBY)
+      require "logtail"
+      logger = Logtail::Logger.new(Logtail::LogDevices::HTTP.new("token", flush_interval: 60, #{ingest.device_options}))
+      logger.info("parent line")
+      logger.close
+      Process.wait(fork { 3.times { |n| logger.info("child line " + n.to_s) } })
+    RUBY
+
+    expect(result.status).to be_success, result.stderr
+    expect(ingest.messages).to contain_exactly("parent line", "child line 0", "child line 1", "child line 2")
+    # Lines written after close are each delivered on their own, see Logtail::LogDevices::HTTP#write
+    expect(ingest.batch_sizes).to eq([1, 3])
+  end
+
+  it "delivers a child's lines when it calls flush before leaving with exit!, which skips at_exit hooks" do
+    result = run_ruby(<<-RUBY)
+      require "logtail"
+      logger = Logtail::Logger.new(Logtail::LogDevices::HTTP.new("token", flush_interval: 60, #{ingest.device_options}))
+      logger.info("parent line")
+      Process.wait(fork do
+        begin
+          logger.info("child line")
+          logger.flush
+        ensure
+          exit!(0) # the way Resque ends a job's process
+        end
+      end)
+    RUBY
+
+    expect(result.status).to be_success, result.stderr
+    expect(ingest.messages).to contain_exactly("parent line", "child line")
+  end
+end
diff --git a/spec/logtail/log_devices/http_spec.rb b/spec/logtail/log_devices/http_spec.rb
index 1d390f5..9adcb17 100755
--- a/spec/logtail/log_devices/http_spec.rb
+++ b/spec/logtail/log_devices/http_spec.rb
@@ -110,6 +110,93 @@
       http.send(:flush)
       http.close
     end
+
+    it "delivers in the calling thread when no outlet thread runs" do
+      messages = []
+      stub = stub_request(:post, "https://in.logs.betterstack.com/").to_return do |request|
+        messages.concat(MessagePack.unpack(Zlib::Inflate.inflate(request.body)).map { |line| line["message"] })
+        { status: 202 }
+      end
+      http = described_class.new("MYKEY", flush_continuously: false)
+      http.write(Logtail::LogEntry.new("INFO", time, nil, "test log message 1", nil, nil))
+      http.write(Logtail::LogEntry.new("INFO", time, nil, "test log message 2", nil, nil))
+
+      http.flush
+
+      expect(stub).to have_been_requested.once
+      expect(messages).to eq(["test log message 1", "test log message 2"])
+      http.close
+    end
+
+    it "doesn't raise when delivering in the calling thread fails, whatever the error" do
+      http = described_class.new("MYKEY", flush_continuously: false)
+      http.write(Logtail::LogEntry.new("INFO", time, nil, "test log message", nil, nil))
+
+      # WebMock refuses to connect with an error that isn't a StandardError
+      expect { http.flush }.not_to raise_error
+      http.close
+    end
+
+    it "doesn't warn about @last_resp when it can't connect while Ruby shuts down, with warnings on" do
+      # Ruby 2.7 and older warn about an instance variable that is read before it's set
+      result = run_ruby(<<-RUBY)
+        $VERBOSE = true
+        require "logtail"
+        # Like Net::HTTP before Ruby 4.0 while Ruby shuts down: it can't start the thread that
+        # times out connecting
+        Net::HTTP.prepend(Module.new { def start(*); raise ThreadError, "can't alloc thread"; end })
+        http = Logtail::LogDevices::HTTP.new("token", flush_continuously: false, ingesting_host: "127.0.0.1", ingesting_port: 1, ingesting_scheme: "http")
+        logger = Logtail::Logger.new(http)
+        logger.info("line")
+        logger.flush
+      RUBY
+
+      expect(result.status).to be_success, result.stderr
+      expect(result.stderr).not_to include("@last_resp not initialized")
+    end
+
+    it "waits about 5 seconds at most for the outlet thread to deliver" do
+      allow_any_instance_of(Net::HTTP).to receive(:request) { sleep } # Better Stack never answers
+      http = described_class.new("MYKEY")
+      http.write(Logtail::LogEntry.new("INFO", time, nil, "test log message", nil, nil))
+
+      flushing = Process.clock_gettime(Process::CLOCK_MONOTONIC)
+      http.flush
+      expect(Process.clock_gettime(Process::CLOCK_MONOTONIC) - flushing).to be_between(4, 7)
+
+      http.instance_variable_get(:@flush_thread).kill.join
+      http.instance_variable_get(:@request_outlet_thread).kill.join
+    end
+
+    it "waits up to 5 seconds for a slow host when it delivers in the calling thread" do
+      # Takes 3 seconds to answer each request
+      slow_ingest = Class.new(LocalIngestServer) do
+        private
+
+        def serve(socket)
+          def socket.write(*)
+            sleep 3
+            super
+          end
+          super
+        end
+      end.new
+      result = run_ruby(<<-RUBY)
+        require "logtail"
+        # A request per line, which flush delivers as no outlet thread runs
+        http = Logtail::LogDevices::HTTP.new("token", flush_continuously: false, batch_size: 1, #{slow_ingest.device_options})
+        logger = Logtail::Logger.new(http)
+        logger.info("first line")
+        logger.info("second line")
+        logger.flush
+      RUBY
+
+      # A request that times out would drop the one after it
+      expect(result.status).to be_success, result.stderr
+      expect(slow_ingest.messages).to contain_exactly("first line", "second line")
+    ensure
+      slow_ingest.stop if slow_ingest
+    end
   end
 
   # Testing a private method because it helps break down our tests
diff --git a/spec/logtail/logger_flush_all_spec.rb b/spec/logtail/logger_flush_all_spec.rb
new file mode 100644
index 0000000..2d698ca
--- /dev/null
+++ b/spec/logtail/logger_flush_all_spec.rb
@@ -0,0 +1,82 @@
+require "spec_helper"
+
+# Rails flushes Rails.logger after every request (Rails::Rack::Logger calls
+# ActiveSupport::LogSubscriber.flush_all!), and Logger#flush returns right away for that call
+# instead of waiting for delivery. These run ActiveSupport's real flush_all! in another process, so
+# an ActiveSupport that calls flush differently fails here. The ingesting host holds its answers
+# until the direct flush, so a flush_all! that waited for delivery would take 5 seconds.
+describe Logtail::Logger, "#flush called by ActiveSupport::LogSubscriber.flush_all!" do
+  {
+    "a Logtail::Logger" => "logger",
+    "a Logtail::Logger in ActiveSupport::TaggedLogging" => "ActiveSupport::TaggedLogging.new(logger)",
+    "a Logtail::Logger in an ActiveSupport::BroadcastLogger, as Rails 7.1 and later wrap it" =>
+      "ActiveSupport::BroadcastLogger.new(logger)",
+    "a Logtail::Logger in ActiveSupport::TaggedLogging in an ActiveSupport::BroadcastLogger" =>
+      "ActiveSupport::BroadcastLogger.new(ActiveSupport::TaggedLogging.new(logger))"
+  }.each do |description, rails_logger|
+    it "returns right away for #{description}, and flushing it directly delivers" do
+      if rails_logger.include?("BroadcastLogger") && Gem.loaded_specs["activesupport"].version < Gem::Version.new("7.1")
+        skip "ActiveSupport::BroadcastLogger was added in ActiveSupport 7.1"
+      end
+
+      result = run_ruby(<<-RUBY)
+        require "logger" # ActiveSupport 6.1 needs it loaded first since concurrent-ruby 1.3.5
+        require "active_support"
+        require "active_support/log_subscriber"
+        require "active_support/tagged_logging"
+        require "json"
+        require "logtail"
+
+        # The ingesting host keeps the lines of every request, but answers only once `answering` is set
+        delivered = []
+        answering = false
+        server = TCPServer.new("127.0.0.1", 0)
+        Thread.new do
+          loop do
+            Thread.new(server.accept) do |socket|
+              while socket.gets
+                length = 0
+                while (header = socket.gets) != "\\r\\n"
+                  length = header.split(":")[1].to_i if header.downcase.start_with?("content-length:")
+                end
+                delivered.concat(MessagePack.unpack(Zlib::Inflate.inflate(socket.read(length))).map { |line| line["message"] })
+                sleep 0.01 until answering
+                socket.write("HTTP/1.1 202 Accepted\\r\\nContent-Length: 0\\r\\n\\r\\n")
+              end
+            end
+          end
+        end
+
+        # Logs a line for every request, like Rails' ActionController::LogSubscriber
+        class RequestLogSubscriber < ActiveSupport::LogSubscriber
+          def request(event)
+            info(event.payload[:message])
+          end
+        end
+        RequestLogSubscriber.attach_to(:app)
+
+        logger = Logtail::Logger.new(Logtail::LogDevices::HTTP.new("token", flush_interval: 60,
+          ingesting_host: "127.0.0.1", ingesting_port: server.addr[1], ingesting_scheme: "http"))
+        ActiveSupport::LogSubscriber.logger = #{rails_logger}
+        ActiveSupport::Notifications.instrument("request.app", message: "Completed 200 OK")
+
+        flushing = Process.clock_gettime(Process::CLOCK_MONOTONIC)
+        ActiveSupport::LogSubscriber.flush_all!
+        flush_all_seconds = Process.clock_gettime(Process::CLOCK_MONOTONIC) - flushing
+        delivered_by_flush_all = delivered.dup
+
+        answering = true
+        ActiveSupport::LogSubscriber.logger.flush
+        puts JSON.generate("flush_all_seconds" => flush_all_seconds, "delivered_by_flush_all" => delivered_by_flush_all,
+          "delivered_by_flush" => delivered.drop(delivered_by_flush_all.size))
+      RUBY
+
+      expect(result.status).to be_success, result.stderr
+      expect(JSON.parse(result.stdout)).to match(
+        "flush_all_seconds" => a_value < 1,
+        "delivered_by_flush_all" => [],
+        "delivered_by_flush" => ["Completed 200 OK"]
+      )
+    end
+  end
+end
diff --git a/spec/logtail/logger_spec.rb b/spec/logtail/logger_spec.rb
index c95312f..f24d6f1 100755
--- a/spec/logtail/logger_spec.rb
+++ b/spec/logtail/logger_spec.rb
@@ -250,6 +250,30 @@
     end
   end
 
+  describe "#flush" do
+    let(:device) { Logtail::LogDevices::HTTP.new("my_source_token") }
+    let(:extra_device) { Logtail::LogDevices::HTTP.new("my_source_token") }
+    let(:logger) { Logtail::Logger.new(device, extra_device) }
+
+    it "flushes the log devices, so that what was logged is delivered" do
+      expect(device).to receive(:flush)
+      expect(extra_device).to receive(:flush)
+      logger.flush
+    end
+
+    it "returns right away when Rails flushes it after every request" do
+      # Stands in for ActiveSupport::LogSubscriber.flush_all!, which calls Rails.logger.flush
+      log_subscriber = Module.new do
+        def self.flush_all!(logger)
+          logger.flush
+        end
+      end
+      expect(device).not_to receive(:flush)
+      expect(extra_device).not_to receive(:flush)
+      log_subscriber.flush_all!(logger)
+    end
+  end
+
   describe "#formatter=" do
     it "should not allow changing the formatter when the device is HTTP" do
       http_device = Logtail::LogDevices::HTTP.new("source_token")
diff --git a/spec/support/processes.rb b/spec/support/processes.rb
index 15332fe..994ef21 100644
--- a/spec/support/processes.rb
+++ b/spec/support/processes.rb
@@ -5,11 +5,13 @@
 require "zlib"
 
 # A local stand-in for the Better Stack ingesting host, for tests that run logtail in a separate
-# process: it answers every request with 202 and records the message of every line it receives.
+# process: it answers every request with 202 and records the message of every line it receives,
+# and how many lines each request held.
 class LocalIngestServer
   def initialize
     @server = TCPServer.new("127.0.0.1", 0)
     @messages = []
+    @batch_sizes = []
     @lock = Mutex.new
     @connections = []
     @thread = Thread.new do
@@ -26,6 +28,10 @@ def messages
     @lock.synchronize { @messages.dup }
   end
 
+  def batch_sizes
+    @lock.synchronize { @batch_sizes.dup }
+  end
+
   def stop
     @thread.kill.join
     @connections.each { |connection| connection.kill.join }
@@ -42,7 +48,10 @@ def serve(socket)
         headers[name.downcase] = value.strip
       end
       lines = MessagePack.unpack(Zlib::Inflate.inflate(socket.read(headers["content-length"].to_i)))
-      @lock.synchronize { @messages.concat(lines.map { |line| line["message"] }) }
+      @lock.synchronize do
+        @messages.concat(lines.map { |line| line["message"] })
+        @batch_sizes << lines.size
+      end
       socket.write("HTTP/1.1 202 Accepted\r\nContent-Length: 0\r\n\r\n")
     end
   rescue IOError, SystemCallError