Skip to main content
Your progress
0 of 5 lessons complete0%
T4 / M4.7 / L1 OF 5 / Engineer TIER / 9 min

RI vs SP vs CUD vs Spot: the four levers

Outcome

By the end of this lesson, you will be able to distinguish the four commitment instruments (RI, SP, CUD, Spot), match workloads to the instrument that fits, and avoid the classic mistake of putting stateful workloads on Spot.


TierEngineer
JTBD”Pick the right commitment instrument for each workload class instead of one-size-fits-all RI purchasing.”
PersonasFinOps Lead · Platform Engineer · Engineering Leader
PrerequisitesT0, Foundations (rack rate vs billing) · M4.1, Maturity
Time9 minutes
Bloom verbDistinguish (Analyze), Match (Apply), Avoid (Evaluate)

1. Concept

The major clouds offer four commitment-style instruments that trade flexibility for discount. Knowing which fits which workload is the foundation of commitment portfolio strategy.

Terminal window
INSTRUMENT PROVIDER SCOPE TYPICAL (3YR NO-UP)
──────────────────────────────────────────────────────────────────────
Reserved Inst AWS Specific instance family ~40% (3-yr no-up)
Azure (similar) Specific SKU ~38% (3-yr no-up)
Savings Plan AWS $/hour spend commitment ~30% (3-yr no-up)
across compute platform
CUD GCP Specific machine type ~57% (3-yr resource)
~20% (1-yr flexible)
Spot All three Stateless workloads; 50-90% (no commit)
interruptible

The figures above are the 3-year no-upfront rates, a common middle choice, not the maxima. All-upfront terms go higher: 3-year RIs reach roughly 60–72%, and 3-year Compute Savings Plans ~50%+, depending on instrument, region, and payment option. Real-world discount depends on the customer’s usage profile and commitment terms.

Selection criteria

The right instrument depends on the workload, not on what “everyone uses”:

Terminal window
WORKLOAD PROFILE BEST INSTRUMENT
──────────────────────────────────────────────────────────────────
Steady 24/7 production with known RI (AWS/Azure) or
instance family CUD (GCP) for the floor
Mixed compute usage; instance types Savings Plan (most flexible)
will change over time
Specific family, stable usage, Reserved Instance
no migrations planned (cheapest for known stable)
Stateless batch / dev work Spot
Burst-mode, unpredictable On-demand
(don't commit to unpredictable)
Multi-year predictable 3-year commit (largest discount)
1-year predictable 1-year commit (moderate discount)
Brand-new workload (no history) On-demand for 1-2 quarters,
then commit once profile is known

Savings Plan vs Reserved Instance: the AWS choice

Both are AWS instruments and the distinction matters:

Terminal window
RESERVED INSTANCE (RI):
Commits to a SPECIFIC instance type, AZ, and OS
Locked: if you change instance type, the RI doesn't apply
Pros: slightly cheaper for very stable known workloads (~5% more)
Cons: rigid; doesn't adapt to instance type changes
SAVINGS PLAN (SP):
Commits to $/hour SPEND across the compute platform
Applies to any EC2, Fargate, or Lambda usage
Pros: flexible; moves with workload type changes
Cons: slightly less discount than RI (~5%)
FOR MOST TEAMS, SP IS THE BETTER DEFAULT:
Workload types evolve over time
The 5% extra savings from RI rarely justifies the rigidity
Engineering changes to instance types shouldn't break the commit

A common starting position: 70-80% of the commitment portfolio in SP, 20-30% in RI for specific high-confidence stable workloads.

Spot: the asterisk

Spot is dramatically cheaper but comes with eviction risk:

Terminal window
SPOT ADVANTAGES:
Cheapest possible compute (often 70-90% below on-demand)
No commitment required
Cancellation-free
SPOT TRADE-OFFS:
Cloud can reclaim instance with short notice (2 min on AWS)
Cannot be used for:
- Databases (single-replica or stateful)
- Single-replica services
- Latency-sensitive synchronous workloads
- Long-running stateful jobs that can't checkpoint
Best for:
- Batch jobs (especially stateless or checkpoint-tolerant)
- K8s nodes with workloads that tolerate eviction
- Stateless services with rolling updates
- CI/CD runners
- Data pipelines with retry logic

Spot is the right answer for batch and stateless workloads. The wrong answer for stateful or latency-critical. Misusing Spot creates incident exposure that the savings don’t justify.

The blended portfolio

Most mature orgs end up with a blend across all four:

Terminal window
TYPICAL PORTFOLIO MIX (mature org):
On-demand: 15-25% (unpredictable bursts + new workloads)
Reserved: 10-20% (stable family-specific)
Savings Plan: 35-45% (most workloads)
Spot: 15-25% (batch / stateless)
Net effective discount vs all-on-demand: 25-40%

The mix is per-workload, not org-wide. Production has different commitment needs than dev/test.

Common matching mistakes

Terminal window
MISTAKE FIX
──────────────────────────────────────────────────────────────────
RI for unpredictable workload SP (more flexible) or
on-demand (no commit)
SP for very stable known workload RI (slightly cheaper)
Spot for stateful workload On-demand or RI;
eviction would cause incident
On-demand for known steady production Commit (SP or RI);
paying 30% premium for
flexibility you don't need
3-year commitment on brand-new workload Wait 1-2 quarters; observe;
then commit
RI for every team's workload Build the portfolio mix
"to be safe" based on actual profiles

How ZopNight surfaces the mix

ZopNight’s Cost Flow by Purchase Type layout (M3.8.L2) shows the current portfolio mix. The Commitment Optimizer module (covered in subsequent lessons) recommends specific RIs, SPs, or CUDs based on observed usage patterns.

For each recommendation, ZopNight shows:

  • Estimated annual savings
  • Break-even point (when commitment pays for itself)
  • Workload that justifies the commit
  • Confidence level (high if 6+ months of stable usage)

When to defer commitment purchasing

Terminal window
SCENARIO WAIT
──────────────────────────────────────────────────────────────────
New workload (<3 months history) Yes: observe first
Major architectural change planned Yes: let it settle
Org migration in progress Yes: instance types
will change
Crawl-stage maturity Yes: get tagging,
scheduling first
Cloud provider EDP renegotiation coming Possibly: commit terms
may shift

Commitment is a multi-year decision. The cost of waiting one quarter is much smaller than the cost of committing to a bad portfolio.


2. Demo

A mid-size customer’s commitment decisions across three workloads:

Terminal window
WORKLOAD 1: production order-processor
Profile: 24/7, ~6 instances steady floor, ~10 instances peak
History: 12 months stable on m5.xlarge
DECISION:
Floor (6 instances): 1-yr Savings Plan at ~25% discount
Peak above floor (4 instances): on-demand or Spot
(Spot tolerable here: stateless web tier with retries)
RATIONALE:
SP for known floor (rigid commit would be RI; SP for flex)
Spot for burst (stateless; eviction-tolerant)
Net effective discount: ~30%
WORKLOAD 2: nightly ETL pipeline
Profile: 4 hours/night, batch, stateless, restartable
History: 8 months consistent timing
DECISION: Spot (entirely)
RATIONALE:
Maximum cost benefit
Eviction-tolerant by design
Batch nature fits: restart logic exists
Net effective discount: 70-85%
WORKLOAD 3: production database (single instance, primary)
Profile: 24/7, no spot tolerable, m5.2xlarge stable
History: 18 months
DECISION: 3-year Reserved Instance on m5.2xlarge
RATIONALE:
Never Spot (stateful, would cause incident)
Very stable workload, no plans to change family
3-year RI captures maximum discount
Net effective discount: ~38%
PORTFOLIO TOTALS:
Workload 1: ~30% effective discount on $80K/mo
Workload 2: ~80% effective discount on $20K/mo
Workload 3: ~38% effective discount on $25K/mo
Across all three: ~$54K/mo saved vs all on-demand
Annual: ~$650K saved

The right instrument per workload, not one-size-fits-all.


3. Hands-on (5 min)

For your team’s three biggest workloads:

Terminal window
WORKLOAD 1: __________
Profile (steady / burst / batch / stateful): __________
History stability: months ____
Best instrument: __________
Justification: __________
WORKLOAD 2: __________
Profile: __________
History: __________
Best instrument: __________
Justification: __________
WORKLOAD 3: __________
Profile: __________
History: __________
Best instrument: __________
Justification: __________
CURRENT COMMITMENT MIX (your org):
On-demand: ____%
Reserved: ____%
Savings Plan: ____%
Spot: ____%
GAPS noticed:
□ Stable workloads on on-demand (commit candidates)
□ Stateful workloads on Spot (move to on-demand)
□ Batch/stateless on on-demand (move to Spot)
□ Single huge RI portfolio (consider SP for flexibility)

4. Knowledge check

Q1

A workload that is 24/7, stateless web tier with traffic spikes. The best commitment mix:

A. RI for everything
B. Savings Plan for the steady floor + Spot for the burst capacity (stateless tolerates eviction; can restart). The floor is predictable, the burst isn’t. SP captures the floor discount; Spot captures the burst discount.
C. Spot for everything
D. On-demand for everything

Show answer

Correct: B. SP for floor + Spot for burst. Mixed instruments for mixed workload shapes.

Q2

A production database. Spot is:

A. Optimal: cheapest compute
B. Inappropriate. Spot can evict the instance with short notice; databases need persistent compute. Production stateful workloads must use on-demand or RI. The cost of an unexpected database eviction far exceeds the Spot savings.
C. Acceptable with retry logic
D. Random

Show answer

Correct: B. Spot for stateless; not for stateful. Spot mismatch creates incident exposure.

Q3

For an org standardizing commitments, SP vs RI:

A. RI always: bigger discount
B. SP for most workloads (more flexible; moves with instance type changes; only ~5% less discount than RI). RI when usage is very stable per-family and no migrations are planned. The ~5% extra discount from RI rarely justifies the rigidity.
C. Both the same
D. Random

Show answer

Correct: B. SP for flexibility. RI for known-stable per-family.


5. Apply

Use Reports → Cost Flow → Purchase Type layout to see your current portfolio mix. The Commitment Optimizer recommends specific instruments based on observed usage profiles. Apply selectively; defer for new workloads.


Glossary terms touched

Reserved Instance · Savings Plan · CUD · Spot instance · Commitment portfolio


Start with the bill.

Foundations takes about five hours. The first lesson is nine minutes.

Open curriculum. No login. No paywall. 237 lessons across 7 courses, three publicly verifiable credentials. Read it on the train, take the exam on a Saturday, list the credential on your résumé Monday.

5h median time to finish Foundations
0 logins, paywalls, or marketing forms
open curriculum, public credential verifier
Multi-cloud automation· Production-ready in 30 min· SOC 2 · ISO 27001· 30% average cloud cost cut· 4 platforms · 1 console· Multi-cloud automation· Production-ready in 30 min· SOC 2 · ISO 27001· 30% average cloud cost cut· 4 platforms · 1 console·