What Happened
On 2026-08-05 AWS announced two product changes relevant to data-heavy serverless processing and ETL observability:
- AWS Lambda now offers scalable network bandwidth for functions outside a VPC. Bandwidth scales with memory from 625 Mbps at 2 GB up to 3,000 Mbps at 10 GB. The feature applies to functions outside a VPC with >=2 GB memory and must be enabled by requesting the “Network bandwidth per execution environment” quota in AWS Service Quotas. It’s available at no additional charge in all commercial AWS Regions [1].
- AWS Glue Data Quality added a new observation mode for anomaly detection that uses a constant baseline (to reduce false positives for irregular or infrequent arrivals) and improved anomaly predictions. Anomaly detection for Glue ETL jobs is now offered at no additional cost and is available in all commercial regions and AWS GovCloud (US) [2].
Why It Matters to Businesses
- Lower latency and lower per-invocation cost for network-bound Lambdas: For workloads that transfer large volumes (up to multiple TB) and are latency sensitive, higher bandwidth reduces execution time and therefore can reduce cost even if you increase memory to reach the bandwidth thresholds [1].
- Operational simplification for data pipelines: Free anomaly detection for Glue ETL jobs removes a direct product charge for basic data-quality monitoring, making it easier to add automated checks and catch pipeline issues earlier [2].
- Architectural decisions matter: The Lambda bandwidth boost applies only to functions outside a VPC; teams must weigh the trade-off between leaving functions outside a VPC (better network performance) and needing VPC access to databases, caches, or internal APIs.
Kimbodo Engineering Perspective
Both updates remove practical barriers but introduce trade-offs that teams must evaluate.
- Lambda bandwidth vs memory-cost trade-off: Bandwidth scales with memory; increasing function memory to 2–10 GB will increase per-invocation compute price but usually shortens execution time. For sustained heavy network I/O, higher memory can reduce total cost and jitter; for bursty or CPU-bound tasks, the memory bump may not pay off. Measure end-to-end latency and cost-per-request before and after changes.
- VPC vs non-VPC placement: Moving heavy-transfer work outside a VPC to get the bandwidth benefit is viable when the function does not require access to internal resources. If it does, consider alternatives: VPC endpoints, moving the transfer stage to a non-VPC function that stages data to S3, or using container/EC2-based solutions where you control NICs and placement.
- Glue anomaly-mode selection: The constant-baseline observation mode reduces false positives for irregular patterns but is not a universal replacement for trend-aware detectors. Use it for exploratory, ad-hoc, or irregular datasets, and combine modes where you have mixed traffic patterns.
- Operationalization: Enabling features is an account-level change (Lambda quota request) and a pipeline-level configuration (Glue checks). Treat both as release tasks with testing, alerting, and rollback plans.
How We Would Implement It
Lambda: enable and validate bandwidth for heavy network tasks
- Identify candidate functions that are network-bound (large S3 transfers, HTTP uploads/downloads, cross-region copies). Use CloudWatch metrics (duration, network bytes) and distributed traces.
- Ensure candidate functions are deployed outside a VPC. If they currently require VPC access, evaluate moving only the transfer stage out-of-VPC or using S3 as a staging buffer.
- Request the “Network bandwidth per execution environment” quota in AWS Service Quotas for your account/region to enable scalable bandwidth [1].
- Increase function memory to the target tier (>=2 GB up to 10 GB) and run A/B tests comparing duration, error rates and cost-per-request. Monitor with Lambda Insights / CloudWatch and measure network throughput achieved.
- If non-VPC placement is unacceptable, consider alternatives: EC2/containers with tuned NICs, AWS Fargate with ENI controls, S3 multipart uploads with accelerated transfer, or DataSync for bulk transfers.
Glue Data Quality: add anomaly detection into ETL pipelines
- Enable Glue Data Quality checks in your Glue jobs or notebooks and select the new constant-baseline observation mode for datasets with irregular arrivals or exploratory workflows [2].
- Define checks (nulls, value ranges, distribution changes) and set alerting to SNS or EventBridge for failures or anomalies. Persist anomaly history in S3/Glue tables for incident review and ML retraining.
- In production pipelines, combine observation modes: use constant-baseline for noisy/irregular datasets and trend-aware modes for time-series metrics. Add a short A/B validation period to tune thresholds and reduce noisy alerts.
- Automate governance: include tests in CI for Glue job definitions and deploy checks as part of pipeline releases; capture metrics in your existing observability stack.
Risks, Costs and Security
- Cost trade-offs: Increasing Lambda memory raises per-invocation compute price even if total cost may fall due to reduced duration. Run cost models before mass rollout. Glue anomaly detection is free, but Glue job runtime, storage for artifact retention, and monitoring/notification costs still apply [2].
- Security and network posture: Functions outside a VPC have direct internet egress; this can increase exposure and complicate access to internal services. Maintain strict IAM, limit outbound destinations, use TLS, and consider dedicated non-VPC transfer functions that only handle non-sensitive payloads.
- Data governance and privacy: Anomaly telemetry and historical checks may contain PII or sensitive metadata. Treat stored anomaly records as data assets with encryption at rest, access controls, and retention policies.
- Operational risk: Enabling account-level quotas or adding automated anomaly alerts can generate noisy signals if not tuned, causing alert fatigue. Pilot, tune thresholds, and phase rollout to critical pipelines first.
- Feature scope: Lambda bandwidth boost only applies to functions outside a VPC and requires memory >=2 GB; it is not a substitute for network engineering choices like dedicated transfer services or regional placement when determinism and control are required [1].
References: AWS Lambda network bandwidth announcement [1]; AWS Glue Data Quality announcement and pricing update [2].
Where Kimbodo Comes In
Kimbodo builds and operates this in production for businesses — see our AI Application Development practice. Wondering what it would cost for your organization? Get a preliminary range, timeline and architecture in about a minute.
