Skip to main content

SQL Database & Elastic Pool Sizing

Review and act on sizing recommendations for Azure SQL single databases and elastic pools. They appear on the same Sizing page as every other resource type: open Recommendations > Sizing and set the Resource type filter to SQL Databases or SQL Elastic Pools.

All SQL switches are off by default

With every switch off, the page shows only a smaller size in the same tier. Turn on a switch to add cheaper tiers, DTU to vCore, serverless or fewer Hyperscale replicas.

A same-tier size-down with every switch off. The evidence table projects each metric onto the target: P95 and busiest day both stay within limits at half the vCores.

What's Analyzed

SQL sizing is always on; there's nothing to enable. Each scan reads the last 90 days of daily Azure SQL metrics, collected as part of Performance Metrics. A database or pool needs at least 30 days of data before its first recommendation. Azure keeps about 93 days of SQL metrics, so databases that already have history can get recommendations from the first scan.

Coverage

Databases
Single databases on DTU Standard and Premium; vCore General Purpose, Business Critical and Hyperscale on standard-series (Gen5); Hyperscale premium-series and premium-series memory optimized; serverless General Purpose (turn-on-auto-pause only).
Databases not analyzed
master and other system databases, Synapse dedicated SQL pools, free-offer databases, databases in an elastic pool, geo, named and standby secondaries, DTU Basic (nothing smaller exists) and Premium RS, DC-series and Fsv2 hardware, and Hyperscale serverless.
Elastic pools
DTU pools (Basic, Standard, Premium) and vCore pools (General Purpose, Business Critical) on standard-series (Gen5).
Pools not analyzed
Hyperscale pools and empty pools. Premium and Business Critical pools get a recommendation only when In-Memory OLTP storage was recorded and stayed at zero for the whole window.
  • Pools are sized as pools. A database inside an elastic pool is never sized on its own; the pool is the billed unit.
  • Idle databases (no logins or sessions for 30 days and CPU P95 at or under 5%, on a database older than 30 days) get no sizing recommendation. They show in Orphaned Resources instead, as do empty pools. A database quiet for less than 30 days can appear in both. Idle pools that still hold databases are sized.
  • Savings use Azure list prices in USD, like every sizing type, not your billed cost.

SQL Toggles

Toggle Options has a SQL Databases section and a SQL Elastic Pools section. Every switch is off by default. Each one adds a kind of move, and with several on, the page shows the eligible option with the largest saving.

SQL Databases

Tier downgrade
Business Critical to General Purpose, Premium to Standard. The cheaper tiers use slower remote storage and have no read replica.
DTU to vCore
Moves from DTU to vCore sizes. Approximate, so confidence is medium at most.
Serverless
Serverless General Purpose targets, and turning on auto-pause for serverless databases.
Fewer HA replicas
Hyperscale targets with fewer high-availability replicas: a longer failover after a compute failure and less capacity for read-only connections.

SQL Elastic Pools

Tier downgrade
Business Critical to General Purpose, Premium to Standard. Never offered when a database in the pool uses In-Memory OLTP.
DTU to vCore
eDTU pools to vCore pools. Medium confidence at most.

The Recommendation type filter offers Size down, Cheaper tier, DTU to vCore, Serverless and Fewer replicas for databases, and the first three for pools.

Turn on one lever at a time

Start with every switch off for the safest view, then turn on one lever at a time to compare. Every toggle state is in the page URL, so you can share the exact view.

The SQL sections at their defaults, with every switch off.

How a Target Is Chosen

A same-tier size-down is always considered: a smaller DTU objective in the same tier, or fewer vCores on the same hardware (Hyperscale included). Every candidate must pass all of these checks:

  • One step at a time: the target keeps at least 50% of the current DTUs or vCores. The next smaller size is always allowed, so sizes such as S2 and P15 can still step down.
  • P95: projected P95 of CPU, data IO, log IO and instance CPU on the target stays under the aggressiveness ceiling (Conservative 70%, Balanced 80%, Aggressive 90%).
  • Busiest day: projected busiest-day CPU, data IO and log IO stay at or under 100% of the target. This keeps short month-end peaks in the check, which a P95 alone would drop.
  • Hard limits: projected highest workers, sessions, tempdb log and In-Memory OLTP storage stay at or under 60% of the target's limit. At these limits Azure rejects new requests or writes fail.
  • Storage: the current max size fits the target. If it doesn't, but the space in use is at most 80% of the target's max size, the card asks you to lower the max size first. Otherwise there's no recommendation.
  • tempdb: used tempdb is at most 80% of the target's tempdb limit, where the size reports one.
  • In-Memory OLTP: a database that uses it never moves to Standard or General Purpose. A target with a smaller In-Memory OLTP limit needs recorded usage.
  • Zone redundancy: a zone-redundant database never moves to Basic or Standard.
  • Savings: at least $1 a month, with the page's Minimum Savings filter applied on top.

Hard limits ignore aggressiveness

Workers, sessions, tempdb log and In-Memory OLTP storage are checked against the highest value seen in 90 days, at 60% of the target's limit, whatever the aggressiveness setting.

Cross-tier and cross-model targets aren't limited to smaller sizes; the cheapest target that passes every check wins.

Levers

Tier downgrade

Premium to Standard, Business Critical to General Purpose, keeping the license setting. Never from Hyperscale, because Azure allows leaving Hyperscale only within 45 days of migrating in.

DTU to vCore

Standard to General Purpose and Premium to Business Critical, using about 100 Standard or 125 Premium DTUs per vCore. With Tier downgrade also on, Premium can go to General Purpose. Confidence is capped at medium.

DTU to vCore targets are priced license-included

If you use Azure Hybrid Benefit, the real saving is larger.

Data IO on DTU to vCore and cross-tier DTU moves is estimated from Microsoft's per-DTU IOPS ranges. Premium publishes only a floor, so a Premium to Standard move is offered only when data IO is close to zero. This is by design.

Serverless

  • A provisioned General Purpose standard-series database can move to serverless at the same max vCores, with the minimum vCores and auto-pause after 60 minutes.
  • A serverless database with auto-pause turned off gets a recommendation to turn it on. This is a setting change; the database isn't rescaled.
  • Pausing is modeled from the last 30 days of hourly data. An hour counts as quiet only with no logins, zero open sessions and zero CPU, and a missing hour never counts. The database must stay quiet for the 60-minute delay before a paused hour counts.
  • It needs at least 168 observed hours (7 days) in that window and data from the last 48 hours. Active hours are priced at the full max vCores.
  • Never offered to a database in a failover group. Databases only; pools have no serverless lever.

Serverless saves money only when a database is fully quiet for long stretches. It also caps memory at 3 GB per vCore.

Some features keep a database from pausing

Geo-replication, long-term backup retention, a DNS alias on the server, a SQL Data Sync sync database or an elastic jobs job database keep a database awake. StratoLens sees failover groups only, and not right after a database joins one, so check the rest first. Vulnerability assessment scans, viewing auditing or Query Store settings, and tag edits wake a paused database, and the first connection after a pause fails with error 40613 for about a minute. Apps need connection retry logic.

Turning on auto-pause for a serverless database. It's a setting change, and Pausable hours shows how much of the month the database would spend paused.

Fewer HA replicas

Hyperscale only: any lower replica count at the current or a smaller size. A zone-redundant database keeps at least 1 replica. Dropping to 0 replicas sends read scale-out connections to the primary.

Elastic Pools

Pools go through the same checks against the pool's own metrics, plus these rules:

  • The target must hold at least the current number of databases.
  • The per-database max moves to the nearest target setting at or above the current one. If it changes, the card says "Per-database max changes from X to Y."
  • A non-zero per-database min must be valid on the target and fit every database. An eDTU pool with a non-zero per-database min isn't converted to vCore.
  • On DTU targets, the largest database's max size must fit the target's per-database storage limit.
  • Commitment coverage is read for the pool itself, not its databases.

Reading the Card

Selecting a database or pool opens one card in the detail panel. Its title names the kind of move:

Database Size Down Recommendation
Smaller size, same tier and model
Cheaper Tier Recommendation
Tier downgrade
DTU to vCore Recommendation
DTU to vCore, alone or with a tier downgrade
Serverless Recommendation
Serverless move or turning on auto-pause
Fewer HA Replicas Recommendation
Fewer Hyperscale replicas
Pool Size Down Recommendation
Smaller pool, same tier
Cheaper Pool Tier Recommendation
Pool tier downgrade
DTU to vCore Pool Recommendation
eDTU pool to vCore pool

When a database recommendation combines moves, the title follows this order: Serverless, DTU to vCore, Cheaper tier, Fewer HA replicas, Size down. A lever badge lists the moves used, for example "Tier downgrade + DTU to vCore", or Size Change for a plain size-down.

Card sections

Current → Recommended
Configuration, detail lines and monthly cost on each side, for example "Standard S3 (100 DTU)" or "Hyperscale 8 vCores, 2 HA replicas". vCore labels say whether the license is included or uses Azure Hybrid Benefit. Pool details include the per-database max and how many databases the target holds.
Commitment banner
Reservation or savings plan coverage, from the resource's own cost data. Coverage is shown, never subtracted from savings. A reservation banner warns that the reserved vCores go unused unless another database or pool in the same tier and region uses them.
Estimated Savings
Per month, per year and % reduction, at list price in USD.
Recommendation Rationale
Current to target, the projected P95 CPU, data IO and log IO on the target, and the busiest day.
Utilization Evidence
CPU, Data IO, Log IO, Workers and Sessions, each with P95, Busiest day, Target P95 and Target busiest day. Workers and Sessions have only the busiest day. A value that isn't collected shows "-", never 0.
Facts
Instance CPU P95, storage allocated and used, tempdb used (databases), In-Memory OLTP in use or not, and for serverless, Pausable hours per month. Each shows only when the value exists.
Confidence Analysis
Data Sufficiency, Headroom Depth and Workload Stability, each scored out of 100. See Confidence.

Severity is based on the default size-down view: over $200/month is high, over $50 is medium, otherwise low. Savings that need a switch don't raise severity, so a database whose only options need a switch shows as low.

View Performance opens the database or pool in Performance Metrics, and View in Explorer opens it in Resource Explorer. Hide and Unhide work as on other resource types.

Before You Apply a Change

The card's caution list names the risks of that specific move. The ones to plan for:

Every resize drops open connections

Scaling is online, but open connections drop at the final switch, usually for under 30 seconds. Make sure apps retry connections.

Some moves copy the data

The card estimates about 1 minute per GB used, and reverting takes as long. Plan the change in a maintenance window. For databases, these moves take a short, constant time instead: within Standard S0 to S1, within Standard S2 to S12, General Purpose to General Purpose, and within Hyperscale. For pools, moves where both sides are Basic, Standard or General Purpose are constant time; other pool moves are timed from the space used across all its databases.

  • Storage latency: Standard and General Purpose use remote storage (about 5-10 ms) instead of local SSD (about 1-2 ms).
  • Read scale-out: on a move to a tier without a read replica, connections with ApplicationIntent=ReadOnly go to the primary. This doesn't block a recommendation.
  • Max size: when the card says "Lower the max size from X GB to Y GB first", lower the database's max size, then apply the recommendation.

Geo-replicated databases

StratoLens sees failover groups but not other geo-replication. A database in a failover group shows it in the card subtitle. Resize the primary first, then its secondaries. Secondaries aren't included in the savings.

After a Change

  • While Azure is still applying a resize, the database or pool drops off the page. Realized savings are recorded once the resize finishes; StratoLens waits up to 3 days.
  • The recommendation moves to Show Resolved with its realized savings. Days at the old size are left out, and a new recommendation waits for the minimum data window at the new size, so recommendations never stack on old data.
  • Realized savings are tracked for changes to service objective, license, HA replicas, minimum vCores and auto-pause delay on databases, and tier, eDTUs, vCores and license on pools.

Settings

Open Settings > Scanner, expand Sizing Recommendations and find the SQL Databases and SQL Elastic Pools blocks. Each type has its own settings. Like other scanner settings, changes apply on the next scan, and saving needs the Configure Scanner Settings permission.

Defaults

Aggressiveness
Balanced. Sets the P95 ceiling for CPU, data IO, log IO and instance CPU only; the 60% hard-limit ceiling never changes.
Minimum data window (days)
30, range 0 to 90. Databases or pools with fewer days of data get no new recommendations; existing ones keep updating. Below 7 lowers the statistics trust floor, and 0 is a testing mode.

Troubleshooting

A database I expected has no recommendation

Answer

Check, in order: it's a type that isn't analyzed (Basic, pooled, free offer, secondary, system, DC-series or Fsv2); it has fewer than 30 days of data or was resized less than 30 days ago; it's idle (see Orphaned Resources); a resize is still in progress; no smaller target passes the checks (often the busiest day, workers or sessions); the saving is below the Minimum Savings filter; the recommendation is hidden; or a better option needs a switch that's off.

Serverless is never offered

Answer

Serverless needs a provisioned General Purpose standard-series database, not in a failover group, with at least 7 days of hourly data in the last 30 days, data from the last 48 hours and enough fully quiet hours. Business Critical, Hyperscale and DTU databases aren't offered serverless.

Tier downgrade is on but nothing changed

Answer

The database or pool uses In-Memory OLTP, is zone-redundant (never moved to Standard), is Hyperscale, or the cheaper tier fails the IO check. Premium to Standard needs data IO near zero.

The savings don't match my bill

Answer

Savings use Azure list prices in USD. Your bill may include reservations, savings plans, Azure Hybrid Benefit or discounts, and be in another currency.

My recommendation disappeared after I resized

Answer

This is expected. The change is recorded under Show Resolved, and a new recommendation can appear after 30 days at the new size.

Severity is low even though savings are large

Answer

Severity uses the default size-down view. Savings that need a switch don't raise it.

Was this page helpful?

Last updated November 6, 2026