Key takeaways Container logs share the container’s lifecycle. The Kubernetes documentation states that when a pod is evicted from a node, it removes the containers along with their logs. Container lifespans keep shrinking. Sysdig’s 2025 Cloud-Native Security and Usage Report found that 60% of containers live for 60 seconds or less. Collection has to be continuous, not scheduled. A node-level agent (a DaemonSet on Kubernetes) covers every container on a node without changing your applications.

The market has already answered whether telemetry is infrastructure. Cisco paid $28 billion for Splunk. Twelve days apart in 2025, CrowdStrike bought Spain’s Onum and SentinelOne moved to acquire Observo AI, both to embed AI-driven data pipelines directly beneath their platforms, while AI-native pipeline startups keep raising eight-figure rounds on the same thesis. When the largest names in security spend that kind of capital on the layer that collects, filters, and routes data, they are conceding something they rarely say out loud: detection, response, and every AI capability stacked on top are only ever as good as the pipeline underneath.

Offline buffering is a log shipper’s ability to store events in memory or on disk while the destination is unreachable, then forward them automatically once the connection recovers. Without it, every SIEM outage, network partition, or agent restart becomes a permanent gap in your security telemetry. Every SIEM eventually goes down. So does the WAN link to your central collector, and eventually the host running the agent itself. None of these events should cost you log data, but whether they do depends on how your log shipper buffers when it can’t deliver data to the destination.

USB device auditing on Windows means tracking three things: when a device connects, what files move to or from it, and whether your device-control policy stopped what it was supposed to stop. Windows can record all three. Almost none of it is enabled by default. Honeywell’s 2025 Cyber Threat Report counted 1,826 unique USB-borne threats in the first quarter of 2025 alone, including 124 that were completely new. These findings came from Secure Media Exchange scanning stations at industrial sites.

A single Windows Event Collector (WEC) server has three documented scaling limits: disk write speed, concurrent WinRM connections, and registry growth from per-client bookmark data. Microsoft’s current guidance is to plan for roughly 3,000 events per second on average across all subscriptions on commodity hardware. The "10k × 10k" rule of thumb (up to 10,000 clients and 10,000 EPS per collector) still circulates widely. IBM QRadar and Juniper JSA documentation still quote it, both attributing it to Microsoft’s article.

Windows records scheduled task activity in two separate event logs. The Security log holds event IDs 4698-4702, covering task creation, deletion, turning on, turning off, and updates, with the full task definition XML attached. The Microsoft-Windows-TaskScheduler/Operational log holds event IDs 106, 140, and 141 for registration changes, plus 100, 200, and 201 for execution. Windows doesn’t switch on either source by default: Microsoft’s own Tarrask analysis states that "neither of these are audited by default and must be explicitly turned on by an administrator.