MediaLive's Video Aligned Locking lines up redundant pipelines by matching frame content

When the origin switches between unlocked copies, the player shows a jump forward or a repeat ^[7]. "In sports, that is the goal the viewer watched twice," the dev.to write-up on the feature says ^[8]. The cause is ordinary. Jitter on the incoming signal, the moment encoding starts and the points where segments are cut all vary from one encoder to the next ^[5].

Locking is configured in encoderSettings.globalConfiguration, in two layers ^[10]. The mode field, outputLockingMode, takes DISABLED, PIPELINE_LOCKING or EPOCH_LOCKING ^[10]. EPOCH_LOCKING ties both pipelines to the Unix epoch, or to a custom epoch from 2000-01-01T00:00:00 onward, and works only with CMAF Ingest and MediaPackage V2 ^[10]. The method field, outputLockingSettings.pipelineLockingSettings.pipelineLockingMethod, exists only under PIPELINE_LOCKING ^[12]. Set it to VIDEO_ALIGNMENT and each encoder computes a visual signature per frame, then aligns the frames whose content matches ^[12].

The default deserves more attention than the new method. A standard channel ships with PIPELINE_LOCKING, and that mode is best-effort. When it cannot lock, MediaLive keeps processing and does not treat the condition as a fault ^[11]. According to the post, nothing stops and nothing alarms; the failover just stops being frame-accurate ^[11]. Timecode-less SRT and RTMP feeds used to land in approximate synchronization, and the author's gloss on that state was "do not trust the failover" ^[1].

Lock state shows up in a metric. Participating pipelines form a pool with one reference pipeline, whose InputVideoAligned value is always 1. The others report 1 only once they align with it ^[14]. An alarm on the reference cannot fire. Because the channel raises no fault, the non-reference pipeline's InputVideoAligned dropping below 1 is the only signal that failover has lost frame accuracy ^[19]. A dashboard pinned to the reference will show perfect alignment for the life of the channel ^[19].

Linked channels extend the pool to single-pipeline channels, with tight rules. The primary owns the group, the follower points at the primary's ARN in the same account, each primary takes one follower, and the follower works whether or not the primary is running ^[15].

Content matching only holds if both encoders see the same pictures. In a standard channel the upstream sends one signal to both inputs ^[4]. I would want two numbers from a production feed before trusting this in a failover drill: how long a follower takes to report alignment, and how matching behaves on long runs of near-identical frames such as a slate. The write-up does not give either.

The field predates the announcement. VIDEO_ALIGNMENT showed up in the SDK on December 26, 2025, and the What's New post came in September 2026 ^[17]. The gap to the September 12 announcement is about 260 days, so an infrastructure-as-code tool built on the SDK may already expose the field ^[18].