Speaker
Dr
Patrick Fuhrmann
(DESY/dCache.org)
Description
dCache has introduced a new paradigm for scientific storage: storage
events. In traditional interactions there is always the same pattern:
the client requests some operation and the server replies with the
result of that request. If the client wishes to learn the current
status of some resource, it makes a request for this information. To
discover when the state of that resource has changed, a client polls
the service by making periodic requests that query the current status
of the resource. Note that the time between successive queries must
be carefully chosen: querying too often may place considerable strain
on the storage system. Therefore, polling inevitably introduces
latency.
It may happen that it is impossible to avoid placing unacceptable load
on the storage system while providing a sufficiently quick response.
In such situations, a typical solution is for the agent that modifies
the storage system to inform interested parties through some
communication side-channel. This coupling between the client that
interacts with the storage and the agent reacting to the changes
brings several disadvantages: the client must be custom (potentially a
maintenance and platform availability problem), and it must know which
agents it must inform (potentially a discovery problem).
Moreover, some events are apparently spontaneous and are not a result
of client interactions. As an example, consider a client that wishes
to learn when a file currently stored on a high-latency media (such as
tape) is available for reading with low latency. This event is not
triggered as the result of external client interaction, but rather
from the activity of the tape system.
In the storage-event model, clients are alerted to activity within the
storage system that they are interested in. The key distinction is
that this happens without polling: when an interesting event happens,
dCache informs the client with minimal delay. This allows the client
to learn of changes in near realtime, without placing load on the
server.
Storage events are widely applicable. They allow more robust and
scalable solutions to several existing infrastructure challenges. For
example, file catalogues may use storage events to learn of changes
autonomously, without requiring custom clients; clients can request a
large number of files be staged back from tape, and start processing
as soon as files become available.
Storage events also allow for complex and innovative solutions. By
breaking down the connection between the agent modifying storage and
the agent reacting to those changes, new services are possible. For
example, it is possible to couple computational workflows with data
ingest events; domain-specific portals can react to data ingest
quickly, so are always up-to-date with the available data.
dCache provides two mechanisms that support storage events: Kafka and
Server-Sent Events (SSE). These provide complementary solutions that
target different use-cases. Kafka storage-events are more suited for
site-level integration, where the institute running dCache wishes to
augment its behaviour by integrating it with other services. SSE, a
standard web protocol for notification, gives ordinary dCache users
access to storage events, allowing ad-hoc innovation.
In this paper, we will present a review of the storage events concept,
an update on the current support for storage event support in dCache,
and an overview of how various projects are using or plan to use
storage events.
Primary authors
Dr
Albert Rossi
(FNAL)
Dr
Dmitry Litvintsev
(FNAL)
Mr
Jürgen Starek
(DESY)
Ms
Marina Sahakyan
(DESY)
Dr
Olufemi Adeyemi
(DESY)
Dr
Patrick Fuhrmann
(DESY/dCache.org)
Dr
Paul Millar
(DESY)
Mr
Tigran Mkrtchyan
(DESY)
