M-Lab Experiment Developer Responsibilities, Requirements, and Guidelines
This document outlines the responsibilities, requirements, and best practice guidelines for experiment developers operating an experiment on the M-Lab platform. Our team is available to consult with experiment developers on both new and existing experiments.
Experiment Application & Review Process
Measurement Lab (M-Lab) is an open platform for research, and welcomes the involvement of new participants. Experiment Developers must first apply to host their tool on the M-Lab platform. Experiments are vetted by the Experiment Design and Review Committee for their potential usefulness for Internet users and research, differentiation from other experiments, and other factors prior to approval and deployment. The Experiment Design and Review Committee and/or the M-Lab team will provide feedback to the Experiment Developer, and may ask for modifications to be made to ensure the experiment meets all requirements and expectations.
In the interest of managing the use of resources and steering the mission of the project, any experiment running on M-Lab must adhere to the following underlying principles, unless waived by the Experiment Design and Review Committee due to unusual circumstances:
- The experiment being used must be open-source. The software running the experiment must be released under an OSI-approved license. This ensures that its source code and methodology can be examined by interested parties, and that the client-side tools can be integrated into a variety of applications. Each experiment must include client as well as server code, all of which must be licensed as open source. Only the original experiment, when deployed on the platform initially, must openly license client-side code as a reference implementation. The client-side reference implementation can be separate from the version actually deployed to users, as long as it provides a reference for the application-layer protocol deployed. The purpose of an open source client implementation is to allow third parties to more easily integrate the code into their own custom clients. The source code to the tools must be released prior to it being deployed to the M-Lab production platform and remain available to the public while hosted on the M-Lab platform.
- The experiment must produce server-side data. We expect that all experiments produce server-side data intended for public use. It is this public data that makes the work of deploying, maintaining, and running of the experiments valuable to M-Lab and the world.
- The data must be released into the public domain. All data produced by the experiment running on the M-Lab platform must be released into the public domain (using a Creative Commons Zero (CC0) license) and made available for free in a portable, machine readable format (e.g. JSON, JSONL) so that it may be examined and freely used for research purposes for the foreseeable future. The data must be released immediately upon its collection. There is one exception for new experiments. New experiments may embargo data for the shorter of the two periods:
- until a research paper is published utilizing data from the experiment as run on the M-Lab platform or
- one year from the time of the experiment’s initial launch.
After this initial optional embargo period the exception is over, and afterward all data must be released immediately.
- The measurement must be client-initiated. End users of the M-Lab platform must explicitly “opt-in” to running a measurement.
- Experiments must only perform “active measurements”. Here “active” means that the client and server generate the network traffic that is measured. In contrast, “passive” measurements involve observations of user data unrelated to the experiment. For example, an active measurement could send text (e.g. random, synthetic, or prerecorded) to measure download speed, while a passive measurement might record a user’s actual DNS lookups. In all cases, any transmitted data should be absent of any human subject information. M-Lab is for measuring the network not the user.
Experiments must actively seek to minimize the collection of PII on the M-Lab platform, ideally collecting none at all. M-Lab seeks to measure the Internet infrastructure without tracking individual Internet users. M-Lab is committed to preserving user privacy, and experiments running on the platform must not collect any information on M-Lab which M-Lab is not permitted to publish. As such, any PII that is collected requires explicit user consent, as well as that the PII be required for the user’s understanding of the data (that it be collected for a legitimate purpose). For example, the NDT experiment requires that users consent to the saving of their public IP address and no other data is required to be submitted to M-Lab to perform the test. Because NDT results don’t make sense without the IP address, collection of this single instance of PII is required for the test results to make sense at all. Each experiment is different, and each experiment’s data needs are different, but the principle is the same: strict data minimization is required for all data collected, saved, and published by the M-Lab platform. This minimization should restrict the data saved to the subset that is required for the test user to understand their own results, rather than the (much larger) set of all data that could potentially be of future interest to an experimenter. For most experiments, we anticipate that the only PII required by the experiment should be the public-facing IP address of the client.
Many useful experiments require additional information which might include richer PII of interest to the experimenter (e.g. precise geolocation). In order to use M-Lab these experiments must segregate PII and non-PII data; they must use their own infrastructure for any data that might contain PII other than IP addresses; they must not use M-Lab servers or data collection to process PII (other than the IP addresses) in any way; they must provide public documentation of their experiment’s compliance with these requirements to be published on or referenced on the M-Lab website; and they must keep the Experiment Design and Review Committee informed and up to date about all of these practices.
- Experiments are subject to operational auditing and tracing. For all client connections initiated with experiments running on M-Lab servers, M-Lab will implement operational auditing and tracing of the connection. This includes packet capture (.pcap), tcp_info, and other tools used to observe the experiment behavior both in testing and production environments. Furthermore, all data resulting from this operational auditing and tracing is published openly in the M-Lab data archive.
When experimenters leverage the M-Lab ETL pipeline to publish raw test data and/or BigQuery datasets, the experimenter consents to joint controllership of their experiment’s data, allowing M-Lab the legal basis for processing it and making it available publicly on the experimenter’s behalf. In this case, experimenters should reflect their roles as joint controllers on their own privacy policies, communicated at the time a person runs their test, which should clearly describe which data serves what purpose and what legal base is applied for each.
An experiment which fits these criteria may apply by submitting the following information to the Experiment Design and Review Committee:
- A description of the organization building and sponsoring the experiment, if applicable, and an introduction to the team that will be maintaining the experiment’s code and documentation supporting its users.
- A narrative description of the value of hosting the experiment long term with M-Lab.
- A narrative description of the experiment’s research goals and how those goals align with M-Lab’s research goals. What do you hope to study by conducting this experiment? Over what period of time do you intend that this experiment run?
- A description of the experiment’s methodology.
- This should include an in-depth description of the data the experiment gathers and how it is gathered, as well as what information is transmitted in each direction (client to server, server to client). It should make clear how the experiment intends to fit within the principles described above and should additionally describe how the experiment differs from existing M-Lab experiments.
- To the best of your abilities, please estimate the bandwidth needed for your experiment to give M-Lab a sense of the resources needed by the experiment.
- Note: experiments should be novel for M-Lab, both in the data they collect and the experience they offer to the client relative to existing measurement tools available on the M-Lab platform. If the experiment proposed already exists and is open-source then links to access the source code should be included. If research papers or articles already exist describing the experiment, links to publicly available papers, articles, and other documentation should be included.
- An example of the data files generated by the experiment.
- A statement that the Experiment Developers submitting the experiment application have read and understood the M-Lab Conceptual and Technical Scope and Policies document, and will operate the experiment in accordance with their responsibilities as laid out in that document.
- A timeline for deploying the experiment, including what milestones exist and what effort is being put into meeting those milestones.
- Documentation proving that the experiment has been approved or exemption by an Institutional Review Board (IRB), Human Subjects Research Board, or other equivalent reviewing entity as appropriate. As a platform that supports open science, M-Lab is concerned with ensuring that all experiments and services running on its platform infrastructure be compliant with accepted standards for human subjects research. Additionally, because all data resulting from experiments running on the M-Lab platform infrastructure is published in a publicly available data archive, all published data must not include Personally Identifiable Information (PII). To these ends, each Experiment Developer is required to ensure that their experiment code and the data it produces has been reviewed and approved for use in human subjects research by their host organization’s Institutional Review Board (IRB). If the host organization does not have an IRB, there are firms that can provide a human subjects review independently that M-Lab can provide contacts and referrals to. All experimenters must show that their experiment and the resulting data it produces do not produce significant risk to persons choosing to run the experiment, and that PII is not present in the resulting data.
- When experimenters leverage the M-Lab ETL pipeline to publish raw test data and/or BigQuery datasets, the experimenter must consent to M-Lab being a joint controller of the resulting data collected by the experiment by completing and signing the M-Lab Data Processing and Joint Controller Agreement in the M-Lab/Experiment Developer Memorandum of Understanding.
These materials should be submitted to M-Lab. The team will review for clarity and questions around operational capacity and need before submitting to the Experiment Design & Review Committee. Once the experiment application has been submitted to the Experiment Design and Review Committee, they will discuss the application at their next official meeting or via email, and decide whether or not to support the new experiment. The Experiment Design and Review Committee will consider applications based on their adherence to the principles outlined in this document, the unique value add of the experiment, and benefit to the Internet Measurement Community at large. The Committee will share their decision and reasoning with the experiment developers and the M-Lab team.
M-Lab uses docker for virtualization and kubernetes for deployment. Your experiment must work inside a docker container. M-Lab expects a dockerfile which will build your container from source and provide all expected functionality for your experiment and for operational monitoring by the M-Lab team.
Experiments must write test data to a specified directory
Experiments should accept the directory name as a command line parameter or use an environment variable to determine where resulting test data should be stored. Data written to that directory will be automatically uploaded, archived and published to Google Cloud Storage (GCS) by M-Lab.
Every running experiment will have a directory for spooling new measurement data for upload. The name of this directory should be configurable. M-Lab will deploy the experiment configured to save data to some subdirectory of
/var/spool. To upload experiment results to M-Lab’s public archive, simply save the data file(s) to that directory. Immediately after a program calls
close() on a file in that directory, an event listener reads the file and prepares its contents for upload. The uploader will delete the files after successful upload. The current code for the system listener that pushes data to M-Lab’s data archive can be found at: https://github.com/m-lab/pusher.
Experiment Developers should refer to the M-Lab Pusher Design Doc to ensure experiment code meets all of M-Lab’s essential design constraints and follows best practices. Essentially, pusher requires that the files it uploads be:
- Only closed once. Once a file is closed after writing (or moved to the directory), it should never be opened for writing again. Pusher assumes ownership of the file after the file is closed, and Pusher reads files immediately after taking ownership of them.
- Not needed by the system after being closed. Once a file is closed (or moved to the directory), Pusher assumes ownership of the file, and that includes assuming that it is no longer needed and may be deleted after upload.
- Not too old and not too gradually updated. A tunable parameter for Pusher is an age limit. By default it is two hours. If the last-modified time of a file is more than two hours in the past, Pusher assumes that the file’s CLOSE_WRITE event was missed (because delivery of those events is not guaranteed) and that the file is eligible for upload. Files that linger open for more than two hours without being modified therefore run the risk of being uploaded prematurely.
Experiment Operating Environment
Container life-cycle managed by M-Lab staff
To ensure that M-Lab staff can audit software running on the production platform, apply emergency updates if necessary, and maintain production deployment SLOs, M-Lab staff manage the complete container lifecycle. Experiment Developers are required to provide the M-Lab team with access to the experiment source code (which must be openly licensed) and the build process must be automated. The M-Lab team will work with Experiment Developers to test and confirm that experiment code functions as expected, and that all monitoring and other operational requirements are met. M-Lab provides support to experiment developers to test their docker container(s) on testing infrastructure to ensure correct and expected operations. M-Lab will work with Experiment Developers to iterate on issues until the experiment can be deployed safely to production infrastructure.
Access to production machines and containers
Experiment Developers will not have direct access to production servers or individual containers running in the production kubernetes cluster. Experimenters will develop using local docker containers and the M-Lab team will work with Experiment Developers to test working containers on non-production testing infrastructure.
Containers are allocated dedicated public IP addresses
Experiments are allocated one IPv4 address and one IPv6 address per machine (if IPv6 is available at that site). A container application that opens a listening socket in the container will result in a listener on the public IPs on that same port. To achieve the same effect when developing locally, provide the
--net=host argument to your docker run command.
Resource Allocation per experiment
The M-Lab platform is finite: a finite number of sites, operational machines at each site, available capacity per machine, and switch uplink capacity. So, experiments are assigned “resource allocations” that limit the maximum resource utilization over some short time period.
Default Resource Allocations:
- CPU: 1 core
- RAM: 1GB
- Disk space: 20GB
- Disk I/O: 10%
- Network bandwidth: 100 Mbps Avg over 5 min
Limited CPU and Network bandwidth are the most likely to interfere with measurements. So, care must be taken to pace clients in such a way to respect resource utilization limits. Based on the available capacity and experiment requirements, default allocations may be adjusted. If you need more, let us know. We are always interested in expanding available platform resources for all experiments through third-party donations.
These limits safeguard the platform and all other experiments on the platform that operate within resource limits.
Experiment Best Practices
Code version control & unit tests
For development, we strongly recommend:
- A publicly accessible, openly licensed source repository, and
- Unit tests that run every time your code is compiled.
To aid in development and satisfy these recommendations, M-Lab currently uses:
- github.com for our code repositories,
- travis-ci.com and Google Cloud Build to automatically build our code and run our unit tests (integrated with Github), and
- coveralls.io for tracking test coverage of our code (integrated with both GitHub and Travis).
Updating experiment container images
- M-Lab makes a clone of your experiment repository on Github, and automates building of container images for repository pushes and tags.
- If you need to update your experiment, create a PR from the experiment’s upstream repository against M-Lab’s clone, and then notify M-Lab of the PR. To make a PR from an upstream repository against a clone of that repository requires comparing across forks on Github.
- M-Lab will review the PR and, when it is ready, merge it, after which M-Lab’s clone will be tagged, creating a versioned container image which M-Lab operators will deploy through its sandbox cluster, then to the staging cluster, and eventually into production.
Label experiment data with socket UUID
Experiment data should, to the maximum degree possible, be joinable with other data. To aid in this, we recommend that you use the UUID library to label TCP connection data saved by your experiment. The label may be the file name, or additional metadata saved with experiment data. Within a standard M-Lab environment, collecting the UUID for a TCP connection is as simple as:
uuid, err := uuid.FromTCPConn(conn)
If your experiment creates multiple TCP connections per test, collect all UUIDs associated with the test and save them together.
The socket UUID is recorded automatically by the traceroute-caller and tcp-info sidecar services bundled with most experiments. When your experiment includes the UUID of tests, then analysts (including you!) will be able to easily join the experiment data with sidecar service data for richer analysis.
If your experiment is written in a language that does not currently have support in the UUID library, let us know and we will write a library for you to use.
Making your data available in BigQuery
We encourage all experiment developers to write a parser to convert the individual data files generated by your experiment into JSON objects as part of http://github.com/m-lab/etl . This will allow M-Lab to provide BigQuery access to your parsed data in BigQuery. The simplest example parser is the Switch DISCO dataset parser because archive files are JSONL.
Monitoring support for your experiment
The M-Lab team maintains all aspects of deploying and managing M-Lab servers, deploying and managing the lifecycle of experiments provided by Experiment Developers, and are responsible for core operating system and package updates, security, etc. of the platform infrastructure. M-Lab also maintains a robust monitoring system of all platform infrastructure, and health checks for experiment code running in virtual containers platform-wide. Experiment Developers are expected to provide monitoring support within their code, including:
- a livenessProbe command
- a simple counter that increments for every experiment run on each container
- metric labels indicating whether an experiment succeeded or failed
- any additional metrics as needed to allow effective monitoring and support SLOs, allowing M-Lab to more easily monitor the health of your service
The “Four Golden Signals” are the most helpful for creating an SLO for your service:
- Traffic - e.g. number of requests, requests / sec, bytes / sec.
- Errors - e.g. requests that fail, for any reason.
- Test duration - e.g. time to service a request, seconds / request.
- Saturation - e.g. system capacity, a fraction or percentage.
Experiment Support & M-Lab Community Expectations
Experiment Developers should provide links to documentation of their experiment to support end users, as well as a support or contact email address where users may be directed by M-Lab support staff. The M-Lab team will also coordinate with Experiment Developers to place relevant documentation on the measurementlab.net website if desired.
Third-party developers are frequently interested in integrating M-Lab hosted tests into their applications. If this is of interest, then Experiment Developers need to include documentation about how to approach developing with the experiment.
The M-Lab team also maintains community lists/groups/forums where Experiment Developers may engage with and support users directly. We encourage Experiment Developers to be involved in M-Lab community forums.
Experiment Lifecycle Expectations
M-Lab expects Experiment Developers to maintain and support experiment code running on the platform for at least the period of their MoU with M-Lab. Further, M-Lab expects experiments running on the platform to be actively maintained and supported for the term of the MoU.
In the event that an Experiment Developer wishes to retire an experiment, M-Lab will expect to coordinate on timing and communications to users with as much advance notice as possible. Typically decommissioning will be scheduled during the MoU renewal process if an Experiment Developer indicates a desire to discontinue hosting. However, decommissioning an experiment outside of this cycle is possible on a case by case basis.
The M-Lab team reserves the right to halt an experiment out of cycle in the event that experiment:
- is overwhelming M-Lab pod switches, uplinks, or other infrastructure
- is the cause of any security related incident affects the performance of other hosted experiments
- or other reasons not outlined here, but which may necessitate administrative action to ensure our SLOs are met
In the event of such incidents as described above, M-Lab’s first action will be to halt experiment containers and communicate with experiment developers on an appropriate course of action. If an experiment developer is unresponsive, in such an incident, or otherwise, this may be cause for M-Lab to consider permanent decommissioning of the experiment.