Operational Considerations for Streaming Media

1.1. Notes for Contributors and Reviewers

Note to RFC Editor: Please remove this section and its subsections before publication.¶

This section is to provide references to make it easier to review the development and discussion on the draft so far.¶

3.1. Scaling Requirements for Media Delivery

3.2. Path Bandwidth Constraints

Even when the bandwidth requirements for video streams along a path are well understood, additional analysis is required to understand the contraints on bandwidth at various points in the network. This analysis is necessary because media servers may react to bandwith constraints using two independent feedback loops:¶

• Media servers often respond to application-level feedback from the media player that indicates a bottleneck link somewhere along the path, by adjusting the amount of media that the media server will send to the media player in a given timeframe. This is described in greater detail in Section 5.¶
• Media servers also typically implement transport protocols with capacity-seeking congestion controllers that probe for bandwidth, and adjust the sending rate based on transport mechanisms. This is described in greater detail in Section 6.¶

The result is that these two (potentially competing) "helpful" mechanisms each respond to the same bottleneck with no coordination between themselves, so that each is unaware of actions taken by the other, and this can result in QOE for users that is significantly lower than what could have been achieved.¶

In one example, if a media server overestimates the available bandwidth to the media player,¶

• the transport protocol detects loss due to congestion, and reduces its sending window size per round trip,¶
• the media server adapts to application-level feedback from the media player, and reduces its own sending rate,¶
• the transport protocol sends media at the new, lower rate, and confirms that this new, lower rate is "safe", because no transport-level loss is occuring, but¶
• because the media server continues to send at the new, lower rate, the transport protocol's maximum sending rate is now limited by the amount of information the media server queues for transmission, so¶
• the transport protocol can't probe for available path bandwidth by sending at a higher rate.¶

In order to avoid these types of situations, which can potentially affect all the users whose streaming media traverses a bottleneck link, there are several possible mitigations that streaming operators can use, but the first step toward mitigating a problem is knowing when that problem occurs.¶

3.3. Path Requirements

The bitrate requirements in Section 3.1 are per end-user actively consuming a media feed, so in the worst case, the bitrate demands can be multiplied by the number of simultaneous users to find the bandwidth requirements for a router on the delivery path with that number of users downstream. For example, at a node with 10,000 downstream users simultaneously consuming video streams, approximately 80 Gbps might be necessary in order for all of them to get typical content at 1080p resolution.¶

However, when there is some overlap in the feeds being consumed by end users, it is sometimes possible to reduce the bandwidth provisioning requirements for the network by performing some kind of replication within the network. This can be achieved via object caching with delivery of replicated objects over individual connections, and/or by packet-level replication using multicast.¶

To the extent that replication of popular content can be performed, bandwidth requirements at peering or ingest points can be reduced to as low as a per-feed requirement instead of a per-user requirement.¶

3.4. Caching Systems

When demand for content is relatively predictable, and especially when that content is relatively static, caching content close to requesters, and pre-loading caches to respond quickly to initial requests is often useful (for example, HTTP/1.1 caching is described in [I-D.ietf-httpbis-cache]). This is subject to the usual considerations for caching - for example, how much data must be cached to make a significant difference to the requester, and how the benefits of caching and pre-loading caches balances against the costs of tracking "stale" content in caches and refreshing that content.¶

It is worth noting that not all high-demand content is "live" content. One relevant example is when popular streaming content can be staged close to a significant number of requesters, as can happen when a new episode of a popular show is released. This content may be largely stable, so low-cost to maintain in multiple places throughout the Internet. This can reduce demands for high end-to-end bandwidth without having to use mechanisms like multicast.¶

Caching and pre-loading can also reduce exposure to peering point congestion, since less traffic crosses the peering point exchanges if the caches are placed in peer networks, especially when the content can be pre-loaded during off-peak hours, and especially if the transfer can make use of "Lower-Effort Per-Hop Behavior (LE PHB) for Differentiated Services" [RFC8622], "Low Extra Delay Background Transport (LEDBAT)" [RFC6817], or similar mechanisms.¶

All of this depends, of course, on the ability of a content provider to predict usage and provision bandwidth, caching, and other mechanisms to meet the needs of users. In some cases (Section 3.5), this is relatively routine, but in other cases, it is more difficult (Section 3.6, Section 3.7).¶

And as with other parts of the ecosystem, new technology brings new challenges. For example, with the emergence of ultra-low-latency streaming, responses have to start streaming to the end user while still being transmitted to the cache, and while the cache does not yet know the size of the object. Some of the popular caching systems were designed around cache footprint and had deeply ingrained assumptions about knowing the size of objects that are being stored, so the change in design requirements in long-established systems caused some errors in production. Incidents occurred where a transmission error in the connection from the upstream source to the cache could result in the cache holding a truncated segment and transmitting it to the end user's device. In this case, players rendering the stream often had the video freeze until the player was reset. In some cases the truncated object was even cached that way and served later to other players as well, causing continued stalls at the same spot in the video for all players playing the segment delivered from that cache node.¶

3.5. Predictable Usage Profiles

Historical data shows that users consume more videos and at a higher bit rate than they did in the past on their connected devices. Improvements in the codecs that help with reducing the encoding bitrates with better compression algorithms could not have offset the increase in the demand for the higher quality video (higher resolution, higher frame rate, better color gamut, better dynamic range, etc.). In particular, mobile data usage has shown a large jump over the years due to increased consumption of entertainment as well as conversational video.¶

3.6. Unpredictable Usage Profiles

Although TCP/IP has been used with a number of widely used applications that have symmetric bandwidth requirements (similar bandwidth requirements in each direction between endpoints), many widely-used Internet applications operate in client-server roles, with asymmetric bandwidth requirements. A common example might be an HTTP GET operation, where a client sends a relatively small HTTP GET request for a resource to an HTTP server, and often receives a significantly larger response carrying the requested resource. When HTTP is commonly used to stream movie-length video, the ratio between response size and request size can become arbitrarily large.¶

For this reason, operators may pay more attention to downstream bandwidth utilization when planning and managing capacity. In addition, operators have been able to deploy access networks for end users using underlying technologies that are inherently asymmetric, favoring downstream bandwidth (e.g. ADSL, cellular technologies, most IEEE 802.11 variants), assuming that users will need less upstream bandwidth than downstream bandwidth. This strategy usually works, except when it fails because application bandwidth usage patterns have changed in ways that were not predicted.¶

One example of this type of change was when peer-to-peer file sharing applications gained popularity in the early 2000s. To take one well-documented case ([RFC5594]), the Bittorrent application created "swarms" of hosts, uploading and downloading files to each other, rather than communicating with a server. Bittorrent favored peers who uploaded as much as they downloaded, so that new Bittorrent users had an incentive to significantly increase their upstream bandwidth utilization.¶

The combination of the large volume of "torrents" and the peer-to-peer characteristic of swarm transfers meant that end user hosts were suddenly uploading higher volumes of traffic to more destinations than was the case before Bittorrent. This caused at least one large Internet service provider (ISP) to attempt to "throttle" these transfers in order to to mitigate the load that these hosts placed on their network. These efforts were met by increased use of encryption in Bittorrent, and complaints to regulators calling for regulatory action.¶

The BitTorrent case study is just one example, but the example is included here to make it clear that unpredicted and unpredictable massive traffic spikes may not be the result of natural disasters, but they can still have significant impacts.¶

Especially as end users increase use of video-based social networking applications, it will be helpful for access network providers to watch for increasing numbers of end users uploading significant amounts of content.¶

3.7. Extremely Unpredictable Usage Profiles

The causes of unpredictable usage described in Section 3.6 were more or less the result of human choices, but we were reminded during a post-IETF 107 meeting that humans are not always in control, and forces of nature can cause enormous fluctuations in traffic patterns.¶

In his talk, Sanjay Mishra [Mishra] reported that after the CoViD-19 pandemic broke out in early 2020,¶

• Comcast's streaming and web video consumption rose by 38%, with their reported peak traffic up 32% overall between March 1 to March 30,¶
• AT&T reported a 28% jump in core network traffic (single day in April, as compared to pre stay-at-home daily average traffic), with video accounting for nearly half of all mobile network traffic, while social networking and web browsing remained the highest percentage (almost a quarter each) of overall mobility traffic, and¶
• Verizon reported similar trends with video traffic up 36% over an average day (pre COVID-19)}.¶

We note that other operators saw similar spikes during this time period. Craig Labowitz [Labovitz] reported¶

• Weekday peak traffic increases over 45%-50% from pre-lockdown levels,¶
• A 30% increase in upstream traffic over their pre-pandemic levels, and¶
• A steady increase in the overall volume of DDoS traffic, with amounts exceeding the pre-pandemic levels by 40%. (He attributed this increase to the significant rise in gaming-related DDoS attacks ([LabovitzDDoS]), as gaming usage also increased.)¶

Subsequently, the Internet Architecture Board (IAB) held a COVID-19 Network Impacts Workshop [IABcovid] in November 2020. Given a larger number of reports and more time to reflect, the following observations from the draft workshop report are worth considering.¶

• Participants describing different types of networks reported different kinds of impacts, but all types of networks saw impacts.¶
• Mobile networks saw traffic reductions and residential networks saw significant increases.¶
• Reported traffic increases from ISPs and Internet Exchange Points (IXP) over just a few weeks were as big as the traffic growth over the course of a typical year, representing a 15-20% surge in growth to land at a new normal that was much higher than anticipated.¶
• At DE-CIX Frankfurt, the world's largest Internet Exchange Point in terms of data throughput, the year 2020 has seen the largest increase in peak traffic within a single year since the IXP was founded in 1995.¶
• The usage pattern changed significantly as work-from-home and videoconferencing usage peaked during normal work hours, which would have typically been off-peak hours with adults at work and children at school. One might expect that the peak would have had more impact on networks if it had happened during typical evening peak hours for video streaming applications.¶
• The increase in daytime bandwidth consumption reflected both significant increases in "essential" applications such as videoconferencing and virtual private networks (VPN), and entertainment applications as people watched videos or played games.¶
• At the IXP level, it was observed that physical link utilization increased. This phenomenon could probably be explained by a higher level of uncacheable traffic such as videoconferencing and VPNs from residential users as they stopped commuting and switched to work-at-home.¶

4.1. Ultra Low-Latency

Ultra low-latency delivery of media is defined here as having a glass-to-glass delay target under one second.¶

Some media content providers aim to achieve this level of latency for live media events. This introduces new challenges relative to less-restricted levels of latency requirements because this latency is the same scale as commonly observed end-to-end network latency variation (for example, due to effects such as bufferbloat ([CoDel]), Wi-Fi error correction, or packet reordering). These effects can make it difficult to achieve this level of latency for the general case, and may require tradeoffs in relatively frequent user-visible media artifacts. However, for controlled environments or targeted networks that provide mitigations against such effects, this level of latency is potentially achievable with the right provisioning.¶

Applications requiring ultra low latency for media delivery are usually tightly constrained on the available choices for media transport technologies, and sometimes may need to operate in controlled environments to reliably achieve their latency and quality goals.¶

Most applications operating over IP networks and requiring latency this low use the Real-time Transport Protocol (RTP) [RFC3550] or WebRTC [RFC8825], which uses RTP for the media transport as well as several other protocols necessary for safe operation in browsers.¶

Worth noting is that many applications for ultra low-latency delivery do not need to scale to more than a few users at a time, which simplifies many delivery considerations relative to other use cases.¶

Recommended reading for applications adopting an RTP-based approach also includes [RFC7656]. For increasing the robustness of the playback by implementing adaptive playout methods, refer to [RFC4733] and [RFC6843].¶

Applications with further-specialized latency requirements are out of scope for this document.¶

4.2. Low-Latency Live

Low-latency live delivery of media is defined here as having a glass-to-glass delay target under 10 seconds.¶

This level of latency is targeted to have a user experience similar to traditional broadcast TV delivery. A frequently cited problem with failing to achieve this level of latency for live sporting events is the user experience failure from having crowds within earshot of one another who react audibly to an important play, or from users who learn of an event in the match via some other channel, for example social media, before it has happened on the screen showing the sporting event.¶

Applications requiring low-latency live media delivery are generally feasible at scale with some restrictions. This typically requires the use of a premium service dedicated to the delivery of live video, and some tradeoffs may be necessary relative to what is feasible in a higher latency service. The tradeoffs may include higher costs, or delivering a lower quality video, or reduced flexibility for adaptive bitrates, or reduced flexibility for available resolutions so that fewer devices can receive an encoding tuned for their display. Low-latency live delivery is also more susceptible to user-visible disruptions due to transient network conditions than higher latency services.¶

Implementation of a low-latency live video service can be achieved with the use of low-latency extensions of HLS (called LL-HLS) [I-D.draft-pantos-hls-rfc8216bis] and DASH (called LL-DASH) [LL-DASH]. These extensions use the Common Media Application Format (CMAF) standard [MPEG-CMAF] that allows the media to be packaged into and transmitted in units smaller than segments, which are called chunks in CMAF language. This way, the latency can be decoupled from the duration of the media segments. Without a CMAF-like packaging, lower latencies can only be achieved by using very short segment durations. However, shorter segments means more frequent intra-coded frames and that is detrimental to video encoding quality. CMAF allows us to still use longer segments (improving encoding quality) without penalizing latency.¶

While a LL-HLS client retrieves each chunk with a separate HTTP GET request, a LL-DASH client uses the chunked transfer encoding feature of the HTTP [CMAF-CTE] which allows the LL-DASH client to fetch all the chunks belonging to a segment with a single GET request. An HTTP server can transmit the CMAF chunks to the LL-DASH client as they arrive from the encoder/packager. A detailed comparison of LL-HLS and LL-DASH is given in [MMSP20].¶

4.3. Non-Low-Latency Live

Non-low-latency live delivery of media is defined here as a live stream that does not have a latency target shorter than 10 seconds.¶

This level of latency is the historically common case for segmented video delivery using HLS [RFC8216] and DASH [MPEG-DASH]. This level of latency is often considered adequate for content like news or pre-recorded content. This level of latency is also sometimes achieved as a fallback state when some part of the delivery system or the client-side players do not have the necessary support for the features necessary to support low-latency live streaming.¶

This level of latency can typically be achieved at scale with commodity CDN services for HTTP(s) delivery, and in some cases the increased time window can allow for production of a wider range of encoding options relative to the requirements for a lower latency service without the need for increasing the hardware footprint, which can allow for wider device interoperability.¶

4.4. On-Demand

On-Demand media streaming refers to playback of pre-recorded media based on a user's action. In some cases on-demand media is produced as a by-product of a live media production, using the same segments as the live event, but freezing the manifest after the live event has finished. In other cases, on-demand media is constructed out of pre-recorded assets with no streaming necessarily involved during the production of the on-demand content.¶

On-demand media generally is not subject to latency concerns, but other timing-related considerations can still be as important or even more important to the user experience than the same considerations with live events. These considerations include the startup time, the stability of the media stream's playback quality, and avoidance of stalls and video artifacts during the playback under all but the most severe network conditions.¶

In some applications, optimizations are available to on-demand video that are not always available to live events, such as pre-loading the first segment for a startup time that doesn't have to wait for a network download to begin.¶

5.1. Overview

A simple model of video playback can be described as a video stream consumer, a buffer, and a transport mechanism that fills the buffer. The consumption rate is fairly static and is represented by the content bitrate. The size of the buffer is also commonly a fixed size. The fill process needs to be at least fast enough to ensure that the buffer is never empty, however it also can have significant complexity when things like personalization or ad workflows are introduced.¶

The challenges in filling the buffer in a timely way fall into two broad categories: 1. content selection and 2. content variation. Content selection comprises all of the steps needed to determine which content variation to offer the client. Content variation is the number of content options that exist at any given selection point. A common example, easily visualized, is Adaptive BitRate (ABR), described in more detail below. The mechanism used to select the bitrate is part of the content selection, and the content variation are all of the different bitrate renditions.¶

ABR is a sort of application-level response strategy in which the streaming client attempts to detect the available bandwidth of the network path by observing the successful application-layer download speed, then chooses a bitrate for each of the video, audio, subtitles and metadata (among the limited number of available options) that fits within that bandwidth, typically adjusting as changes in available bandwidth occur in the network or changes in capabilities occur during the playback (such as available memory, CPU, display size, etc.).¶

5.2. Adaptive Encoding

Media servers can provide media streams at various bitrates because the media has been encoded at various bitrates. This is a so-called "ladder" of bitrates, that can be offered to media players as part of the manifest that describes the media being requested by the media player, so that the media player can select among the available bitrate choices.¶

The media server may also choose to alter which bitrates are made available to players by adding or removing bitrate options from the ladder delivered to the player in subsequent manifests built and sent to the player. This way, both the player, through its selection of bitrate to request from the manifest, and the server, through its construction of the bitrates offered in the manifest, are able to affect network utilization.¶

5.3. Adaptive Segmented Delivery

ABR playback is commonly implemented by streaming clients using HLS [RFC8216] or DASH [MPEG-DASH] to perform a reliable segmented delivery of media over HTTP. Different implementations use different strategies [ABRSurvey], often relying on proprietary algorithms (called rate adaptation or bitrate selection algorithms) to perform available bandwidth estimation/prediction and the bitrate selection.¶

Many systems will do an initial probe or a very simple throughput speed test at the start of a video playback. This is done to get a rough sense of the highest video bitrate in the ABR ladder that the network between the server and player will likely be able to provide under initial network conditions. After the initial testing, clients tend to rely upon passive network observations and will make use of player side statistics such as buffer fill rates to monitor and respond to changing network conditions.¶

The choice of bitrate occurs within the context of optimizing for one or more metrics monitored by the client, such as highest achievable video quality or lowest chances for a rebuffering event (playback stall).¶

5.4. Advertising

A variety of business models exist for producers of streaming media. Some content providers derive the majority of the revenue associated with streaming media directly from consumer subscriptions or one-time purchases. Others derive the majority of their streaming media associated revenue from advertising. Many content providers derive income from a mix of these and other sources of funding. The inclusion of advertising alongside or interspersed with streaming media content is therefore common in today's media landscape.¶

Some commonly used forms of advertising can introduce potential user experience issues for a media stream. This section provides a very brief overview of a complex and evolving space, but a complete coverage of the potential issues is out of scope for this document.¶

The same techniques used to allow a media player to switch between renditions of different bitrates at segment or chunk boundaries can also be used to enable the dynamic insertion of advertisements (herafter referred to as "ads").¶

Ads may be inserted either with Client Side Ad Insertion (CSAI) or Server Side Ad Insertion (SSAI). In CSAI, the ABR manifest will generally include links to an external ad server for some segments of the media stream, while in SSAI the server will remain the same during advertisements, but will include media segments that contain the advertising. In SSAI, the media segments may or may not be sourced from an external ad server like with CSAI.¶

In general, the more targeted the ad request is, the more requests the ad service needs to be able to handle concurrently. If connectivity is poor to the ad service, this can cause rebuffering even if the underlying video assets (both content and ads) are able to be accessed quickly. The less targeted, the more likely the ad requests can be consolidated and can leverage the same caching techniques as the video content.¶

In some cases, especially with SSAI, advertising space in a stream is reserved for a specific advertiser and can be integrated with the video so that the segments share the same encoding properties such as bitrate, dynamic range, and resolution. However, in many cases ad servers integrate with a Supply Side Platform (SSP) that offers advertising space in real-time auctions via an Ad Exchange, with bids for the advertising space coming from Demand Side Platforms (DSPs) that collect money from advertisers for delivering the advertisements. Most such Ad Exchanges use application-level protocol specifications published by the Interactive Advertising Bureau [IAB-ADS], an industry trade organization.¶

This ecosystem balances several competing objectives, and integrating with it naively can produce surprising user experience results. For example, ad server provisioning and/or the bitrate of the ad segments might be different from that of the main video, either of which can sometimes result in video stalls. For another example, since the inserted ads are often produced independently they might have a different base volume level than the main video, which can make for a jarring user experience.¶

Additionally, this market historically has had incidents of ad fraud (misreporting of ad delivery to end users for financial gain). As a mitigation for concerns driven by those incidents, some SSPs have required the use of players with features like reporting of ad delivery, or providing information that can be used for user tracking. Some of these and other measures have raised privacy concerns for end users.¶

In general this is a rapidly developing space with many considerations, and media streaming operators engaged in advertising may need to research these and other concerns to find solutions that meet their user experience, user privacy, and financial goals. For further reading on mitigations, [BAP] has published some standards and best practices based on user experience research.¶

5.5. Bitrate Detection Challenges

This kind of bandwidth-measurement system can experience trouble in several ways that are affected by networking and transport protocol issues. Because adaptive application-level response strategies are often using rates as observed by the application layer, there are sometimes inscrutable transport-level protocol behaviors that can produce surprising measurement values when the application-level feedback loop is interacting with a transport-level feedback loop.¶

A few specific examples of surprising phenomena that affect bitrate detection measurements are described in the following subsections. As these examples will demonstrate, it is common to encounter cases that can deliver application level measurements that are too low, too high, and (possibly) correct but varying more quickly than a lab-tested selection algorithm might expect.¶

These effects and others that cause transport behavior to diverge from lab modeling can sometimes have a significant impact on bitrate selection and on user QOE, especially where players use naive measurement strategies and selection algorithms that don't account for the likelihood of bandwidth measurements that diverge from the true path capacity.¶

5.6. Measurement Collection

Media players use measurements to guide their segment-by-segment adaptive streaming requests, but may also provide measurements to streaming media providers.¶

In turn, providers may base analytics on these measurements, to guide decisions such as whether adaptive encoding bitrates in use are the best ones to provide to media players, or whether current media content caching is providing the best experience for viewers.¶

To that effect, the Consumer Technology Association (CTA) who owns the Web Application Video Ecosystem (WAVE) project has published two important specifications.¶

• CTA-2066: Streaming Quality of Experience Events, Properties and Metrics¶

[CTA-2066] specifies a set of media player events, properties, QOE metrics and associated terminology for representing streaming media QOE across systems, media players and analytics vendors. While all these events, properties, metrics and associated terminology is used across a number of proprietary analytics and measurement solutions, they were used in slightly (or vastly) different ways that led to interoperability issues. CTA-2066 attempts to address this issue by defining a common terminology as well as how each metric should be computed for consistent reporting.¶

• CTA-5004: Common Media Client Data (CMCD)¶

Many assume that the CDNs have a holistic view into the health and performance of the streaming clients. However, this is not the case. The CDNs produce millions of log lines per second across hundreds of thousands of clients and they have no concept of a "session" as a client would have, so CDNs are decoupled from the metrics the clients generate and report. A CDN cannot tell which request belongs to which playback session, the duration of any media object, the bitrate, or whether any of the clients have stalled and are rebuffering or are about to stall and will rebuffer. The consequence of this decoupling is that a CDN cannot prioritize delivery for when the client needs it most, prefetch content, or trigger alerts when the network itself may be underperforming. One approach to couple the CDN to the playback sessions is for the clients to communicate standardized media-relevant information to the CDNs while they are fetching data. [CTA-5004] was developed exactly for this purpose.¶

6.1. UDP and Its Behavior

For most of the history of the Internet, we have trusted UDP-based applications to limit their impact on other users. One of the strategies used was to use UDP for simple query-response application protocols, such as DNS, which is often used to send a single-packet request to look up the IP address for a DNS name, and return a single-packet response containing the IP address. Although it is possible to saturate a path between a DNS client and DNS server with DNS requests, in practice, that was rare enough that DNS included few mechanisms to resolve contention between DNS users and other users (whether they are also using DNS, or using other application protocols that share the same pathways).¶

In recent times, the usage of UDP-based applications that were not simple query-response protocols has grown substantially, and since UDP does not provide any feedback mechanism to senders to help limit impacts on other users, application-level protocols such as RTP [RFC3550] have been responsible for the decisions that TCP-based applications have delegated to TCP - what to send, how much to send, and when to send it. Because UDP itself has no transport-layer feedback mechanisms, UDP-based applications that send and receive substantial amounts of information are expected to provide their own feedback mechanisms, and to respond to the feedback the application receives. This expectation is most recently codified in Best Current Practice [RFC8085].¶

In contrast to adaptive segmented delivery over a reliable tansport as described in Section 5.3, some applications deliver streaming media using an unreliable transport, and rely on a variety of approaches, including:¶

• raw MPEG Transport Stream ("MPEG-TS")-formatted video [MPEG-TS] over UDP, which makes no attempt to account for reordering or loss in the transport,¶
• RTP [RFC3550], which can notice loss and repair some limited reordering,¶
• SCTP [RFC4960], which can use partial reliability [RFC3758] to recover from some loss, but can abandon recovery to limit head-of-line blocking, and¶
• SRT [SRT], which can use forward error correction and time-bound retransmission to recover from loss within certain limits, but can abandon recovery to limit head-of-line blocking.¶

Under congestion and loss, approaches like the above generally experiences transient video artifacts more often and delay of playback effects less often, as compared with reliable segment transport. Often one of the key goals of using a UDP-based transport that allows some unreliability is to reduce latency and better support applications like videoconferencing, or for other live-action video with interactive components, such as some sporting events.¶

Congestion avoidance strategies for deployments using unreliable transport protocols vary widely in practice, ranging from being entirely unresponsive to congestion, to using feedback signaling to change encoder settings (as in [RFC5762]), to using fewer enhancement layers (as in [RFC6190]), to using proprietary methods to detect QOE issues and turn off video in order to allow less bandwidth-intensive media such as audio to be delivered.¶

RTP relies on RTCP Sender and Receiver Reports [RFC3550] as its own feedback mechanism, and even includes Circuit Breakers for Unicast RTP Sessions [RFC8083] for situations when normal RTP congestion control has not been able to react sufficiently to RTP flows sending at rates that result in sustained packet loss.¶

The notion of "Circuit Breakers" has also been applied to other UDP applications in [RFC8084], such as tunneling packets over UDP that are potentially not congestion-controlled (for example, "Encapsulating MPLS in UDP", as described in [RFC7510]). If streaming media is carried in tunnels encapsulated in UDP, these media streams may encounter "tripped circuit breakers", with resulting user-visible impacts.¶

6.2. TCP and Its Behavior

For most of the history of the Internet, we have trusted TCP to limit the impact of applications that sent a significant number of packets, in either or both directions, on other users. Although early versions of TCP were not particularly good at limiting this impact [RFC0793], the addition of Slow Start and Congestion Avoidance, as described in [RFC2001], were critical in allowing TCP-based applications to "use as much bandwidth as possible, but to avoid using more bandwidth than was possible". Although dozens of RFCs have been written refining TCP decisions about what to send, how much to send, and when to send it, since 1988 [Jacobson-Karels] the signals available for TCP senders remained unchanged - end-to-end acknowledgements for packets that were successfully sent and received, and packet timeouts for packets that were not.¶

The success of the largely TCP-based Internet is evidence that the mechanisms TCP used to achieve equilibrium quickly, at a point where TCP senders do not interfere with other TCP senders for sustained periods of time, have been largely successful. The Internet continued to work even when the specific mechanisms used to reach equilibrium changed over time. Because TCP provides a common tool to avoid contention, as some TCP-based applications like FTP were largely replaced by other TCP-based applications like HTTP, the transport behavior remained consistent.¶

In recent times, the TCP goal of probing for available bandwidth, and "backing off" when a network path is saturated, has been supplanted by the goal of avoiding growing queues along network paths, which prevent TCP senders from reacting quickly when a network path is saturated. Congestion control mechanisms such as COPA [COPA18] and BBR [I-D.cardwell-iccrg-bbr-congestion-control] make these decisions based on measured path delays, assuming that if the measured path delay is increasing, the sender is injecting packets onto the network path faster than the receiver can accept them, so the sender should adjust its sending rate accordingly.¶

Although TCP behavior has changed over time, the common practice of implementing TCP as part of an operating system kernel has acted to limit how quickly TCP behavior can change. Even with the widespread use of automated operating system update installation on many end-user systems, streaming media providers could have a reasonable expectation that they could understand TCP transport protocol behaviors, and that those behaviors would remain relatively stable in the short term.¶

6.3. QUIC and Its Behavior

The QUIC protocol, developed from a proprietary protocol into an IETF standards-track protocol [RFC9000], turns many of the statements made in Section 6.1 and Section 6.2 on their heads.¶

Although QUIC provides an alternative to the TCP and UDP transport protocols, QUIC is itself encapsulated in UDP. As noted elsewhere in Section 7.1, the QUIC protocol encrypts almost all of its transport parameters, and all of its payload, so any intermediaries that network operators may be using to troubleshoot HTTP streaming media performance issues, perform analytics, or even intercept exchanges in current applications will not work for QUIC-based applications without making changes to their networks. Section 7 describes the implications of media encryption in more detail.¶

While QUIC is designed as a general-purpose transport protocol, and can carry different application-layer protocols, the current standardized mapping is for HTTP/3 [I-D.ietf-quic-http], which describes how QUIC transport features are used for HTTP. The convention is for HTTP/3 to run over UDP port 443 [Port443] but this is not a strict requirement.¶

When HTTP/3 is encapsulated in QUIC, which is then encapsulated in UDP, streaming operators (and network operators) might see UDP traffic patterns that are similar to HTTP(S) over TCP. Since earlier versions of HTTP(S) rely on TCP, UDP ports may be blocked for any port numbers that are not commonly used, such as UDP 53 for DNS. Even when UDP ports are not blocked and HTTP/3 can flow, streaming operators (and network operators) may severely rate-limit this traffic because they do not expect to see legitimate high-bandwidth traffic such as streaming media over the UDP ports that HTTP/3 is using.¶

As noted in Section 5.5.2, because TCP provides a reliable, in-order delivery service for applications, any packet loss for a TCP connection causes "head-of-line blocking", so that no TCP segments arriving after a packet is lost will be delivered to the receiving application until the lost packet is retransmitted, allowing in-order delivery to the application to continue. As described in [RFC9000], QUIC connections can carry multiple streams, and when packet losses do occur, only the streams carried in the lost packet are delayed.¶

A QUIC extension currently being specified ([I-D.ietf-quic-datagram]) adds the capability for "unreliable" delivery, similar to the service provided by UDP, but these datagrams are still subject to the QUIC connection's congestion controller, providing some transport-level congestion avoidance measures, which UDP does not.¶

As noted in Section 6.2, there is an increasing interest in transport protocol behaviors that respond to delay measurements, instead of responding to packet loss. These behaviors may deliver improved user experience, but in some cases have not responded to sustained packet loss, which exhausts available buffers along the end-to-end path that may affect other users sharing that path. The QUIC protocol provides a set of congestion control hooks that can be used for algorithm agility, and [RFC9002] defines a basic algorithm with transport behavior that is roughly similar to TCP NewReno [RFC6582]. However, QUIC senders can and do unilaterally choose to use different algorithms such as loss-based CUBIC [RFC8312], delay-based COPA or BBR, or even something completely different.¶

The Internet community does have experience with deploying new congestion controllers without melting the Internet. As noted in [RFC8312], both the CUBIC congestion controller and its predecessor BIC have significantly different behavior from Reno-style congestion controllers such as TCP NewReno [RFC6582], but both CUBIC and BIC were added to the Linux kernel in order to allow experimentation and analysis, and both were then selected as the default TCP congestion controllers in Linux, and both were deployed globally.¶

The point mentioned in Section 6.2 about TCP congestion controllers being implemented in operating system kernels is different with QUIC. Although QUIC can be implemented in operating system kernels, one of the design goals when this work was chartered was "QUIC is expected to support rapid, distributed development and testing of features", and to meet this expectation, many implementers have chosen to implement QUIC in user space, outside the operating system kernel, and to even distribute QUIC libraries with their own applications. It is worth noting that streaming operators using HTTP/3, carried over QUIC, can expect more frequent deployment of new congestion controller behavior than has been the case with HTTP/1 and HTTP/2, carried over TCP.¶

It is worth considering that if TCP-based HTTP traffic and UDP-based HTTP/3 traffic are allowed to enter operator networks on roughly equal terms, questions of fairness and contention will be heavily dependent on interactions between the congestion controllers in use for TCP-based HTTP traffic and UDP-based HTTP/3 traffic.¶

7.1. General Considerations for Media Encryption

The use of strong encryption does provide confidentiality for encrypted streaming media, from the sender to either an intermediary or the ultimate media consumer, and this does prevent Deep Packet Inspection by any intermediary that does not possess credentials allowing decryption. However, even encrypted content streams may be vulnerable to traffic analysis. An intermediary that can identify an encrypted media stream without decrypting it, may be able to "fingerprint" the encrypted media stream of known content, and then match the targeted media stream against the fingerprints of known content. This protection can be lessened if a media provider is repeatedly encrypting the same content. [CODASPY17] is an example of what is possible when identifying HTTPS-protected videos over TCP transport, based either on the length of entire resources being transferred, or on characteristic packet patterns at the beginning of a resource being transferred.¶

If traffic analysis is successful at identifying encrypted content and associating it with specific users, this breaks privacy as certainly as examining decrypted traffic.¶

Because HTTPS has historically layered HTTP on top of TLS, which is in turn layered on top of TCP, intermediaries do have access to unencrypted TCP-level transport information, such as retransmissions, and some carriers exploited this information in attempts to improve transport-layer performance [RFC3135]. The most recent standardized version of HTTPS, HTTP/3 [I-D.ietf-quic-http], uses the QUIC protocol [RFC9000] as its transport layer. QUIC relies on the TLS 1.3 initial handshake [RFC8446] only for key exchange [RFC9001], and encrypts almost all transport parameters itself, with the exception of a few invariant header fields. In the QUIC short header, the only transport-level parameter which is sent "in the clear" is the Destination Connection ID [RFC8999], and even in the QUIC long header, the only transport-level parameters sent "in the clear" are the Version, Destination Connection ID, and Source Connection ID. For these reasons, HTTP/3 is significantly more "opaque" than HTTPS with HTTP/1 or HTTP/2.¶

[I-D.ietf-quic-manageability] discusses manageability of the QUIC transport protocol that is used to encapsulate HTTP/3, focusing on the implications of QUIC's design and wire image on network operations involving QUIC traffic. It discusses what network operators can consider in some detail.¶

More broadly, RFC 9065 [RFC9065], "Considerations around Transport Header Confidentiality, Network Operations, and the Evolution of Internet Transport Protocols" describes the impact of increased encryption of transport headers in general terms.¶

7.2. Considerations for "Hop-by-Hop" Media Encryption

Although the IETF has put considerable emphasis on end-to-end streaming media encryption, there are still important use cases that require the insertion of intermediaries.¶

There are a variety of ways to involve intermediaries, and some are much more intrusive than others.¶

From a content provider's perspective, a number of considerations are in play. The first question is likely whether the content provider intends that intermediaries are explicitly addressed from endpoints, or whether the content provider is willing to allow intermediaries to "intercept" streaming content transparently, with no awareness or permission from either endpoint.¶

If a content provider does not actively work to avoid interception by intermediaries, the effect will be indistinguishable from "impersonation attacks", and endpoints cannot be assumed of any level of privacy.¶

Assuming that a content provider does intend to allow intermediaries to participate in content streaming, and does intend to provide some level of privacy for endpoints, there are a number of possible tools, either already available or still being specified. These include¶

• Server And Network assisted DASH [MPEG-DASH-SAND] - this specification introduces explicit messaging between DASH clients and network elements or between various network elements for the purpose of improving the efficiency of streaming sessions by providing information about real-time operational characteristics of networks, servers, proxies, caches, CDNs, as well as DASH client's performance and status.¶
• "Double Encryption Procedures for the Secure Real-Time Transport Protocol (SRTP)" [RFC8723] - this specification provides a cryptographic transform for the Secure Real-time Transport Protocol that provides both hop-by-hop and end-to-end security guarantees.¶
• Secure Media Frames [SFRAME] - [RFC8723] is closely tied to SRTP, and this close association impeded widespread deployment, because it could not be used for the most common media content delivery mechanisms. A more recent proposal, Secure Media Frames [SFRAME], also provides both hop-by-hop and end-to-end security guarantees, but can be used with other transport protocols beyond SRTP.¶

The choice of whether to involve intermediaries sometimes requires careful consideration. As an example, when ABR manifests were commonly sent unencrypted some networks would modify manifests during peak hours by removing high-bitrate renditions in order to prevent players from choosing those renditions, thus reducing the overall bandwidth consumed for delivering these media streams and thereby improving the network load and the user experience for their customers. Now that ubiquitous encryption typically prevents this kind of modification, in order to maintain the same level of network health and user experience across networks whose users would have benefitted from this intervention a media streaming operator sometimes needs to choose between adding intermediaries who are authorized to change the manifests or adding significant extra complexity to their service.¶

Some resources that might inform other similar considerations are further discussed in [RFC8824] (for WebRTC) and [I-D.ietf-quic-manageability] (for HTTP/3 and QUIC).¶

7.3. Considerations for "End-to-End" Media Encryption

"End-to-end" media encryption offers the potential of providing privacy for streaming media consumers, with the idea being that if an unauthorized intermediary can't decrypt streaming media, the intermediary can't use Deep Packet Inspection to examine HTTP request and response headers and identify the media content being streamed.¶

"End-to-end" media encryption has become much more widespread in the years since the IETF issued "Pervasive Monitoring Is an Attack" [RFC7258] as a Best Current Practice, describing pervasive monitoring as a much greater threat than previously appreciated. After the Snowden disclosures, many content providers made the decision to use HTTPS protection - HTTP over TLS - for most or all content being delivered as a routine practice, rather than in exceptional cases for content that was considered "sensitive".¶

Unfortunately, as noted in [RFC7258], there is no way to prevent pervasive monitoring by an "attacker", while allowing monitoring by a more benign entity who "only" wants to use DPI to examine HTTP requests and responses in order to provide a better user experience. If a modern encrypted transport protocol is used for end-to-end media encryption, intermediary streaming operators are unable to examine transport and application protocol behavior. As described in Section 7.2, only an intermediary streaming operator who is explicitly authorized to examine packet payloads, rather than intercepting packets and examining them without authorization, can continue these practices.¶

[RFC7258] said that "The IETF will strive to produce specifications that mitigate pervasive monitoring attacks", so streaming operators should expect the IETF's direction toward preventing unauthorized monitoring of IETF protocols to continue for the forseeable future.¶