AWS re:Invent 2025: What to Adopt Now

AWS re:Invent 2025 delivered a lot of noise and a handful of signals. If you run production workloads or lead a platform team, the primary keyword you care about is AWS re:Invent 2025 because it sets your next two quarters of technical debt—either paying it down with better primitives or compounding it by waiting too long. The event ran December 1–5, 2025 in Las Vegas, and several launches matured enough to use immediately.(reinvent.awsevents.com)

The shortlist: what’s real, useful, and ready

Three announcements stood out for teams that ship: Lambda durable functions, S3 Vectors, and Graviton5-backed EC2 instances. Each solves a problem most teams already have—reliable multi-step workflows, vector-scale data for AI features, and better price–performance without re-architecting.(aws.amazon.com)

Lambda durable functions: stateful flows without the sprawl

Lambda durable functions bring long-running, multi-step coordination into the Lambda programming model. You can checkpoint progress, pause for up to a year, and resume without paying for idle compute—no external state machine to stitch together, no custom compensating logic for every edge case. Initial general availability landed December 2, 2025, with Python 3.13/3.14 and Node.js 22/24 support, and region coverage expanded again on December 18.(aws.amazon.com)

When to use it: order workflows, user onboarding, human-in-the-loop approvals, and AI agent handoffs where steps span minutes to days. When not to: high-fan-out, massively parallel orchestration with dozens of branches—AWS Step Functions still shines there.

Amazon S3 Vectors: vector storage where your data already lives

S3 Vectors is now generally available (December 2, 2025) with support for up to two billion vectors per index and latencies around 100 ms for frequent queries—forty times the preview scale. It plugs into Bedrock Knowledge Bases and OpenSearch, giving you a native place to park embeddings without babysitting a separate vector database. The pitch: durability, lower cost, and simpler ops for RAG and AI agents.(aws.amazon.com)

If you’ve been juggling OpenSearch or a third-party vector DB for semantic features while storing the rest of the corpus in S3, this reduces cognitive overhead and cost. You’ll still want a performance tier (OpenSearch or similar) for ultra-low-latency, high-QPS workloads—and you should model your hot/cold strategy accordingly.(aws.amazon.com)

Graviton5: real headroom without rewriting

Graviton5 debuted with new M9g instances and claims up to 25% higher performance than the previous generation, with 192 cores and a larger cache helping steady tail latencies. If your fleet is still x86, the bar to try ARM just dropped again; if you’re already on Graviton3/4, expect incremental but meaningful gains for CPU-bound services.(aboutamazon.com)

AWS re:Invent 2025 in context: what matters for builders

Let’s get practical. The most common question after re:Invent is “What do we adopt now, what do we test in a sandbox, and what do we just watch?” Here’s the framework we use in our cloud modernization work:

Adopt / Experiment / Watch framework

Adopt now when the service is GA, the API is stable, and the migration blast radius is controlled. Experiment when the feature is GA but the operational model is new to your team, or it’s in select Regions. Watch when it’s preview-only or when it demands a net-new competency your team doesn’t yet have.

• Adopt: Lambda durable functions for multi-step flows that used to sprawl across queues, tables, and cron jobs. Start with one well-bounded workflow—refunds, document signing, or subscription cancellations.(aws.amazon.com)
• Adopt: S3 Vectors for RAG and similarity search where data gravity is already in S3 and you don’t need sub-50 ms tail latencies. Model costs with vector bucket + index counts.(aws.amazon.com)
• Experiment: Graviton5 on a canary slice of web/API workloads. Validate performance-per-dollar on your code with real traffic.(aboutamazon.com)
• Watch: Agent frameworks and “frontier agents” from the keynotes—promising, but tune for your governance and safety model first.(aboutamazon.com)

How to roll out Lambda durable functions safely

Here’s the thing: orchestration bugs rarely show up in unit tests; they show up as stranded state at 2 a.m. Durable functions reduce the surface area, but you still need discipline. Use this six-step starter plan:

1. Choose a bounded workflow. Pick one with 5–12 steps, existing retries, and painful manual recovery.
2. Map steps and guards. Turn each step into a durable function segment; define idempotency keys and compaction points.
3. Define failure contracts. For every external call, write down how you detect partial success and how you resume.
4. Use deterministic inputs. Persist request/response snapshots at step boundaries so you can replay.
5. Add human checkpoints. Model “wait” steps with clear SLAs and escalation paths.
6. Back-test with production traces. Re-run last 30 days of events through a shadow durable function and compare outputs.

Launch tip: start in Regions that already have durable functions and your data. AWS added 14 Regions on December 18, 2025, but check quotas and KMS integration before you scale.(aws.amazon.com)

People also ask: Is Step Functions dead?

No. Step Functions still fits complex fan-out/fan-in, visual orchestration, and cross-service governance. Durable functions shine when you want orchestration without leaving the Lambda mental model.

Designing with S3 Vectors (without over-optimizing on day one)

Don’t think of S3 Vectors as “just a cheaper vector DB.” It’s vector storage where most of your unstructured data already is. Use it to reduce hops, simplify lineage, and lower spend on cold and warm embeddings while keeping a smaller hot tier for latency-sensitive paths.(aws.amazon.com)

Practical checklist: your first vector bucket

• Cardinality: estimate vectors per asset type (documents, images, events). Round up; schema drift happens.
• Indexing: plan 1–3 indexes per use case (search, recommendations, anomaly detection). Avoid one giant index “just in case.”
• Metadata: cap at 20–30 meaningful keys per vector; don’t turn metadata into your real database.
• Latency tiers: target 80–150 ms in S3 Vectors; push ultra-hot queries to OpenSearch or memory stores.(aws.amazon.com)
• Security: choose SSE-KMS with per-index keys for multi-tenant apps; audit who can create indexes.
• Lifecycle: set retention policies for embeddings; re-embed on model upgrades, not every release.

People also ask: Is S3 Vectors a drop-in replacement for my vector DB?

In many cases, no. Treat it as your durable, scalable source of vector truth plus a cost-effective warm tier. Keep a smaller, high-throughput engine for the 5–10% of queries that demand it.(aws.amazon.com)

Graviton5 migration playbook (no drama edition)

Moving to ARM isn’t “just flip an AMI,” but it’s not 2020-hard either. Here’s a pragmatic path we’ve used on dozens of services:

1. Inventory: identify CPU-bound services and those with pure interpreted runtimes (Node, Python) where rebuild risk is low.
2. Build: produce multi-arch images and artifacts; for JVM, test JDK 21+ on arm64. For native deps, ensure alpine/musl or glibc variants exist.
3. Benchmark: compare p95 and p99, not just mean. Tail latency improvements often pay the bills.
4. Canary: shift 5–10% of traffic to M9g; watch CPU steal, EBS and network throughput, and GC behavior.
5. Cost: measure real perf-per-dollar (including savings plans/commitments) before scaling the blast radius.
6. Roll: move stateless tiers first, then stateful ones with binlog replication or data mirroring.

What to expect: up to 25% perf gains on CPU-heavy services and steadier caches under load; I/O-bound apps may see less. The big win is fleet density—packing more requests per dollar.(aboutamazon.com)

People also ask: Do I need to rewrite for Graviton5?

Usually not. Most modern stacks are architecture-agnostic. The exceptions are native extensions (image processing, crypto, scientific libs). Build multi-arch CI and watch for subtle differences in SIMD and numerical kernels.

AI agents, responsibly: where to start (and what to avoid)

Keynotes were heavy on agentic AI. If agents are on your 2026 roadmap, start by defining guardrails—identity, budget, and policy—before you give an “ops agent” permission to page-prod or an “engineering agent” rights to merge. AWS is pushing AgentCore and a wave of “frontier agents,” but the build/buy calculus hinges on your audit and safety needs.(aboutamazon.com)

Security hygiene still matters more than hype. If you followed our incident playbooks during the React2Shell saga, you know how to patch and prove quickly—apply that muscle to agent stacks too. See our 72‑Hour Patch‑and‑Prove Plan if you need a refresher on fast, verifiable remediation.

Build vs. buy: the boring—but decisive—questions

Before you spin new infra, ask:

• Will this reduce tickets or just move them?
• Will this consolidate systems or create a parallel universe we’ll regret?
• How will we prove cost/perf wins to finance in 30 days?
• What’s our rollback plan?

For many teams, the winning combo today is: Lambda durable functions for orchestration, S3 Vectors for embeddings at rest, and a targeted Graviton5 pilot where compute costs are noisy. That’s a measurable step forward without re-platforming the world.(aws.amazon.com)

Release notes worth bookmarking

If you need receipts and dates for your change board:

• Lambda durable functions announced Dec 2, 2025; expanded to 14 more Regions Dec 18, 2025.(aws.amazon.com)
• S3 Vectors GA posted Dec 2, 2025 (up to 2B vectors per index; ~100 ms frequent queries).(aws.amazon.com)
• Graviton5 and M9g instances unveiled during re:Invent week (AWS cites up to 25% perf uplift and 192 cores).(aboutamazon.com)
• re:Invent dates: Dec 1–5, 2025.(reinvent.awsevents.com)

What to do next (next 30–60 days)

1. Pick one workflow for durable functions and ship it. Set a 2-week timebox and define “done” as successful replay of a month of production traces.
2. Stand up a vector bucket and index in S3 Vectors; move one RAG use case from your current vector store and compare costs/latency.
3. Run a Graviton5 canary on a stateless service with noisy CPU graphs; track p95 and cost per million requests.
4. Write agent guardrails (identity, budget, policy). Even if you’re not deploying agents yet, this becomes your template.
5. Schedule a 60-minute architecture review to delete one service you no longer need because of the above.

How we can help (and what it costs)

If you want a partner to accelerate this with you, our cloud modernization playbooks cover serverless workflows, ARM migrations, and AI data plumbing end to end. See our services overview and browse a few representative portfolio case studies. If you already know the scope, check ballparks on our pricing page and then get in touch for a concise plan.

FAQ: quick answers your team will ask this week

Does Lambda durable functions replace our Step Functions workflows?

No. Use durable functions for linear or lightly branched flows you want to keep inside Lambda. Keep Step Functions for complex branching, visual coordination, and cross-account workflows.(aws.amazon.com)

Is S3 Vectors enough for our search product?

For many internal features and moderate-QPS user features, yes—especially when data already lives in S3. For ultra-low-latency public search at high QPS, pair it with a hot tier like OpenSearch.(aws.amazon.com)

Will Graviton5 break our dependencies?

Most likely not if you already build multi-arch containers. Audit native extensions and build arm64 variants. Pilot on staging with production traffic replay before going fleet-wide.(techradar.com)

Zooming out

AWS re:Invent 2025 wasn’t just AI fireworks. It gave builders three concrete tools to cut toil and spend without risky rewrites. Durable functions reduce orchestration glue. S3 Vectors puts embeddings where your data gravity lives. Graviton5 raises the floor on price–performance. Pick one, land it, prove the win, and fund the next.