diff --git a/spec/std/file_spec.cr b/spec/std/file_spec.cr
index 1c4a2f5383ad..9b3b7bfce6d3 100644
--- a/spec/std/file_spec.cr
+++ b/spec/std/file_spec.cr
@@ -127,6 +127,33 @@ describe "File" do
         end
       end
     end
+
+    it "can append non-blocking to an existing file" do
+      with_tempfile("append-existing.txt") do |path|
+        File.write(path, "hello")
+        File.write(path, " world", mode: "a", blocking: false)
+        File.read(path).should eq("hello world")
+      end
+    end
+
+    it "returns the actual position after non-blocking append" do
+      with_tempfile("delete-file.txt") do |filename|
+        File.write(filename, "hello")
+
+        File.open(filename, "a", blocking: false) do |file|
+          file.tell.should eq(0)
+
+          file.write "12345".to_slice
+          file.tell.should eq(10)
+
+          file.seek(5, IO::Seek::Set)
+          file.write "6789".to_slice
+          file.tell.should eq(14)
+        end
+
+        File.read(filename).should eq("hello123456789")
+      end
+    end
   end
 
   it "reads entire file" do
diff --git a/src/crystal/event_loop/iocp.cr b/src/crystal/event_loop/iocp.cr
index 163a4602e29a..19a64904af92 100644
--- a/src/crystal/event_loop/iocp.cr
+++ b/src/crystal/event_loop/iocp.cr
@@ -228,10 +228,24 @@ class Crystal::EventLoop::IOCP < Crystal::EventLoop
   end
 
   def write(file_descriptor : Crystal::System::FileDescriptor, slice : Bytes) : Int32
-    System::IOCP.overlapped_operation(file_descriptor, "WriteFile", file_descriptor.write_timeout, writing: true) do |overlapped|
+    bytes_written = System::IOCP.overlapped_operation(file_descriptor, "WriteFile", file_descriptor.write_timeout, writing: true) do |overlapped|
+      overlapped.offset = UInt64::MAX if file_descriptor.system_append?
+
       ret = LibC.WriteFile(file_descriptor.windows_handle, slice, slice.size, out byte_count, overlapped)
       {ret, byte_count}
     end.to_i32
+
+    # The overlapped offset forced a write to the end of the file, but unlike
+    # synchronous writes, an asynchronous write incorrectly updates the file
+    # pointer: it merely adds the number of written bytes to the current
+    # position, disregarding that the offset might have changed it.
+    #
+    # We could seek before the async write (it works), but a concurrent fiber or
+    # parallel thread could also seek and we'd end up overwriting instead of
+    # appending; we need both the offset + explicit seek.
+    file_descriptor.system_seek(0, IO::Seek::End) if file_descriptor.system_append?
+
+    bytes_written
   end
 
   def wait_writable(file_descriptor : Crystal::System::FileDescriptor) : Nil
diff --git a/src/crystal/system/win32/file.cr b/src/crystal/system/win32/file.cr
index e4e45ee3a664..c97600a573fc 100644
--- a/src/crystal/system/win32/file.cr
+++ b/src/crystal/system/win32/file.cr
@@ -12,7 +12,7 @@ module Crystal::System::File
   # On Windows we cannot rely on the system mode `FILE_APPEND_DATA` and
   # keep track of append mode explicitly. When writing data, this ensures to only
   # write at the end of the file.
-  @system_append = false
+  getter? system_append = false
 
   def self.open(filename : String, mode : String, perm : Int32 | ::File::Permissions, blocking : Bool?) : FileDescriptor::Handle
     perm = ::File::Permissions.new(perm) if perm.is_a? Int32
diff --git a/src/crystal/system/win32/file_descriptor.cr b/src/crystal/system/win32/file_descriptor.cr
index 333f3c5a617d..b5c91e95e4f7 100644
--- a/src/crystal/system/win32/file_descriptor.cr
+++ b/src/crystal/system/win32/file_descriptor.cr
@@ -18,6 +18,10 @@ module Crystal::System::FileDescriptor
 
   @system_blocking = true
 
+  def system_append?
+    false
+  end
+
   private def system_read(slice : Bytes) : Int32
     handle = windows_handle
     if ConsoleUtils.console?(handle)
@@ -160,7 +164,7 @@ module Crystal::System::FileDescriptor
     FileDescriptor.system_info windows_handle
   end
 
-  private def system_seek(offset, whence : IO::Seek) : Nil
+  def system_seek(offset, whence : IO::Seek) : Nil
     if LibC.SetFilePointerEx(windows_handle, offset, nil, whence) == 0
       raise IO::Error.from_winerror("Unable to seek", target: self)
     end
diff --git a/src/crystal/system/win32/iocp.cr b/src/crystal/system/win32/iocp.cr
index 0c08cfe21bc6..cfb1fd1aef7e 100644
--- a/src/crystal/system/win32/iocp.cr
+++ b/src/crystal/system/win32/iocp.cr
@@ -264,6 +264,11 @@ struct Crystal::System::IOCP
     def initialize(@handle : LibC::HANDLE)
     end
 
+    def offset=(value : UInt64)
+      @overlapped.union.offset.offset = LibC::DWORD.new!(value)
+      @overlapped.union.offset.offsetHigh = LibC::DWORD.new!(value >> 32)
+    end
+
     def wait_for_result(timeout, & : WinError ->)
       wait_for_completion(timeout)