Deployment Considerations

Sarracenia data pumps are often placed in network designs near demarcation points, to provide an application level demarcation point to allow for security scanning and to limit visibility into different zones. Pumps may either have all services incorporated on a single server, or commonly the main broker is on a dedicated node, and the nodes that do the data transfer are called Transport Engines. In high performance deployments, each transport engine may have a local broker to deal with local transfers and transformations.

Transport Engines

Transport engines are the data servers queried by subscribers, by the end users, or other pumps. The subscribers read the notices and fetch the corresponding data, using the indicated protocol. The software to serve the data can be either SFTP or HTTP (or HTTPS.) For specifics of configuring the servers for use, please consult the documentation of the servers themselves. Also note that additional protocols can be enabled through the use of do_ plugins, as described in the Programming Guide.

IPv6

A sample pump was implemented on a small VPS with IPv6 enabled. A client from far away connected to the rabbitmq broker using IPv6, and the subscription to the apache httpd worked without issues. It just works. There is no difference between IPv4 and IPv6 for sarra tools, which are agnostic of IP addresses.

On the other hand, one is expected to use hostnames, since use of IP addresses will break certificate use for securing the transport layer (TLS, aka SSL) No testing of IP addresses in URLs (in either IP version) has been done.

Designs

There are many different arrangements in which sarracenia can be used.

Dataless: where one runs just sarracenia on top of a broker with no local transfer engines. This is used, for example to run sr_winnow on a site to provide redundant data sources.
Standalone: the most obvious one, run the entire stack on a single server, openssh and a web server as well the broker and sarra itself. Makes a complete data pump, but without any redundancy.
Switching/Routing: Where, in order to achieve high performance, a cluster of standalone nodes are placed behind a load balancer. The load balancer algorithm is just round-robin, with no attempt to associate a given source with a given node. This has the effect of pumping different parts of large files through different nodes. So one will see parts of files announced by such pump, to be re-assembled by subscribers.
Data Dissemination: Where in order to serve a large number of clients, multiple identical servers, each with a complete mirror of data
FIXME:: ok, opened big mouth, now need to work through the examples.

Dataless or S=0

A configuration which includes only the AMQP broker. This configuration can be used when users have access to disk space on both ends and only need a mediator. This is the configuration of sftp.science.gc.ca, where the HPC disk space provides the storage so that the pump does not need any, or pumps deployed to provide redundant HA to remote data centres.

Note

FIXME: sample configuration of shovels, and sr_winnow (with output to xpublic) to allow subscribers in the SPC to obtain data from either edm or dor.

Note that while a configuration can be dataless, it can still make use of rabbitmq clustering for high availability requirements (see rabbitmq clustering below.)

Winnowed Dataless

Another example of a dataless pump would be to provide product selection from two upstream sources using sr_winnow. The sr_winnow is fed by shovels from upstream sources, and the local clients just connect to this local pump. sr_winnow takes care of only presenting the products from the first server to make them available. One would configure sr_winnow to output to the xpublic exchange on the pump.

Local subscribers just point at the output of sr_winnow on the local pump. This is how feeds are implemented in the Storm prediction centres of ECCC, where they may download data from whichever national hub produces data first.

Dataless With Sr_poll

The sr_poll program can verify if products on a remote server are ready or modified. For each of the product, it emits an notification message on the local pump. One could use sr_subscribe anywhere, listen to notification messages and get the products (provided the credentials to access it)

Standalone

In a standalone configuration, there is only one node in the configuration. It runs all components and shares none with any other nodes. That means the Broker and data services such as sftp and apache are on the one node.

One appropriate usage would be a small non-24x7 data acquisition setup, to take responsibility of data queuing and transmission away from the instrument. It is restarted when the opportunity arises. It is just a matter of installing and configuring all a data flow engine, a broker, and the package itself on a single server. The ddi systems are generally configured this way.

Switching/Routing

In a switching/routing configuration, there is a pair of machines running a single broker for a pool of transfer engines. So each transfer engine’s view of the file space is local, but the queues are global to the pump.

Note: On such clusters, all nodes that run a component with the same config file created by default have an identical queue_name. Targetting the same broker, it forces the queue to be shared. If it should be avoided, the user can just overwrite the default queue_name inserting ${HOSTNAME}. Each node will have its own queue, only shared by the node instances. ex.: queue_name q_${BROKER_USER}.${PROGRAM}.${CONFIG}.${HOSTNAME} )

Often there is internal traffic of data acquired before it is finally published. As a means of scaling, often transfer engines will also have brokers to handle local traffic, and only publish final products to the main broker. This is how ddsr systems are generally configured.

Security Considerations

This section is meant to provide insight to those who need to perform a security review of the application prior to implementation.

Client

All credentials used by the application are stored in the ~/.config/sarra/credentials.conf file, and that file is forced to 600 permissions.

Server/Broker

Authentication used by transport engines is independent of that used for the brokers. A security assessment of rabbitmq brokers and the various transfer engines in use is needed to evaluate the overall security of a given deployment.

The most secure method of transport is the use of SFTP with keys rather than passwords. Secure storage of sftp keys is covered in documentation of various SSH or SFTP clients. The credentials file just points to those key files.

For Sarracenia itself, password authentication is used to communicate with the AMQP broker, so implementation of encrypted socket transport (SSL/TLS) on all broker traffic is strongly recommended.

Sarracenia users are actually users defined on rabbitmq brokers. Each user Alice, on a broker to which she has access:

can create and publish to any exchange that starts with xs_Alice_

has an exchange xr_Alice, where she reads her report messages.

can configure (read from and acknowledge) queues named qs_Alice_.* to bind to exchanges

Alice can create and destroy her own queues and exchanges, but no-one else’s.

Alice can only post data that she is publishing (it will refer back to her)

Alice can also read (or subscribe to) any exchange whose name ends in public.

Alice can thus create an exchange others can subscribe to with the following name: xs_Alice_public

Alice cannot create any exchanges or other queues not shown above.

Rabbitmq provides the granularity of security to restrict the names of objects, but not their types. Thus, given the ability to create a queue named q_Alice, a malicious Alice could create an exchange named q_Alice_xspecial, and then configure queues to bind to it, and establish a separate usage of the broker unrelated to sarracenia.

To prevent such misuse, sr_audit is a component that is invoked regularly looking for mis-use, and cleaning it up.

Input Validation

Users such as Alice post their notification messages to their own exchange (xs_Alice). Processes which read from user exchanges have a responsibility for validation. The process that reads xs_Alice (likely an sr_sarra) will overwrite any source or cluster heading written into the notification message with the correct values for the current cluster, and the user which posted the notification message.

The checksum algorithm used must also be validated. The algorithm must be known. Similarly, if there is a malformed header of some kind, it should be rejected immediately. Only well-formed notification messages should be forwarded for further processing.

In the case of sr_sarra, the checksum is re-calculated when downloading the data, it ensures it matches the notification message. If they do not match, an error report message is published. If the recompute_checksum option is True, the newly calculated checksum is put into the notification message. Depending on the level of confidence between a pair of pumps, the level of validation may be relaxed to improve performance.

Another difference with inter-pump connections, is that a pump necessarily acts as an agent for all the users on the remote pumps and any other pumps the pump is forwarding for. In that case, the validation constraints are a little different:

source doesn´t matter. (feeders can represent other users, so do not overwrite.)
ensure cluster is not local cluster (as that indicates either a loop or misuse.)

If the notification message fails the non-local cluster test, it should be rejected, and logged (FIME: published … hmm… clarify)

Note

FIXME:

if the source is not good, and the cluster is not good… cannot report back. so just log locally?

Privileged System Access

Ordinary (sources, and subscribers) users operate sarra within their own permissions on the system, like an scp command. The pump administrator account also runs under a normal linux user account and, given requires privileges only on the AMQP broker, but nothing on the underlying operating system. It is convenient to grant the pump administrator sudo privileges for the rabbitmqctl command.

There may be a single task which must operate with privileges: cleaning up the database, which is an easily auditable script that must be run on a regular basis. If all acquisition is done via sarra, then all of the files will belong to the pump administrator, and privileged access is not required for this either.

Glossary

Sarracenia documentation uses a number of words in a particular way. This glossary should make it easier to understand the rest of the documentation.

Source

Someone who wants to ship data to someone else. They do that by advertising a trees of files that are copied from the starting point to one or more pumps in the network. The notification message sources produced tell others exactly where and how to download the files, and Sources have to say where they want the data to go to.

Sources use programs like sr_post.1, sr_watch.1, and sr_poll(1) to create their notification messages.

Subscribers

are those who examine notification messages about files that are available, and download the files they are interested in.

Subscribers use sr_subscribe(1)

Post, Notice, Notification, Advertisement, Announcement

These are AMQP messages build by sr_post, sr_poll, or sr3_watch to let users know that a particular file is ready. The format of these AMQP messages is described by the sr_post(7) manual page. All of these words are used interchangeably. Advertisements at each step preserve the original source of the posting, so that report messages can be routed back to the source.

Report messages

These are AMQP messages (in sr_post(7) format, with _report_ field included) built by consumers of notification messages, to indicate what a given pump or subscriber decided to do with a notification message. They conceptually flow in the opposite direction of notifications in a network, to get back to the source.

Pump or broker

A pump is a host running Sarracenia, a rabbitmq AMQP server (called a broker in AMQP parlance) The pump has administrative users and manage the AMQP broker as a dedicated resource. Some sort of transport engine, like an apache server, or an openssh server, is used to support file transfers. SFTP, and HTTP/HTTPS are the protocols which are fully supported by sarracenia. Pumps copy files from somewhere, and write them locally. They then re-advertise the local copy to its neighbouring pumps, and end user subscribers, they can obtain the data from this pump.

Note

For end users, a pump and a broker is the same thing for all practical purposes. When pump administrators configure multi-host clusters, however, a broker might be running on two hosts, and the same broker could be used by many transport engines. The entire cluster would be considered a pump. So the two words are not always the same.

Dataless Pumps: There are some pumps that have no transport engine, they just mediate transfers for other servers, by making notification messages available to clients and servers in their network area.
Dataless Transfers: Sometimes transfers through pumps are done without using local space on the pump.
Pumping Network: A number of interconnects servers running the sarracenia stack. Each stack determines how it routes items to the next hop, so the entire size or extent of the network may not be known to those who put data into it.
Network Maps: Each pump should provide a network map to advise users of the known destination that they should advertise to send to. FIXME undefined so far.