Many distributed applications deployed on the Internet must
operate continuously with no noticeable interruption of ser-
vice. Such 24/7 availability requirements make the main-
tenance of these application difficult because fixing bugs or
adding new functionality necessitates the online replacement
of the software version by the new one, i.e., a “live update”.
Support for “live update” is therefore essential to allow soft-
ware evolution of critical services. While the problem of
live update has been widely studied and several techniques
have been proposed (e.g., using group communication and
replication), we propose in this paper an original approach
for the dataflow-based programming model (FBP). An in-
teresting property of FBP is its seamless support for multi-
and many-core architectures, which have become the norm
in recent generation of servers and Cloud infrastructures.
We introduce a framework and new algorithms for imple-
menting coordinated non-blocking updates, which do not
only support the replacement of individual software com-
ponents, but also modifications of structural aspects of the
applications independently of the underlying execution in-
frastructure. These algorithms allow us to transparently
orchestrate live updates without halting the executing pro-
gram. We illustrate and evaluate our approach on a web
server application. We present experimental evidence that
our live update algorithms are scalable and have negligible
impact on availability and performance.