i

Description

Sablier de bureau : le temps que l'on attend s'écoule grain par grain.

Source

Suohros · original page

Licence

CC0 Public domain

Original photo, unretouched. The background is a version of it stylised by the site.

Performance and Data Volume

This theme measures how long the user waits (opening, clicking a filter, export) and the product's ability to keep its promise as data grows. It covers…

In brief

This theme measures how long the user waits (opening, clicking a filter, export) and the product's ability to keep its promise as data grows. It covers column cardinality, model size, storage modes (Import, DirectQuery, Dual, Direct Lake), aggregations, number of visuals, and query analysis (Performance Analyzer, DAX Studio, query plans).

On this page

1. 1. Why this theme
2. 2. What the theme measures
3. 3. The 5 levels
4. 4. Milestones by level
5. 5. Detailed reading by level
   1. 🟢 Level 1 — Suffered
   2. 🟢 Level 2 — Measured
   3. 🟡 Level 3 — Managed
   4. 🟡 Level 4 — Budgeted
   5. 🔴 Level 5 — Guaranteed
6. 6. Partial achievement and rating
7. 7. Moving to the next level
8. 8. See also

Level: 🟢→🔴 · Theme code: PRF

1. Why this theme#

A slow report is not “used less”: it is bypassed (Excel export, direct SQL query, local copy), which destroys governance. Commonly accepted usage benchmarks: below about 1 s, the user does not perceive waiting; around 10 s, they disengage (orders of magnitude from ergonomics research; adapt as needed). Performance of an Import model depends first on cardinality and size (VertiPaq columnar engine), then DAX quality (DAX), then the page (number of visuals).

2. What the theme measures#

• Axis: from “it's lagging and we don't know why” to response time guaranteed by a measured pipeline budget.
• Key indicators: page render time (P50/P95), DAX query duration per visual, model size (MB/GB), refresh duration, share of DirectQuery requests > 3 s.
• Units: seconds, MB, number of visuals per page.

3. The 5 levels#

4. Milestones by level#

5. Detailed reading by level#

🟢 Level 1 — Suffered#

• Signs: “it's been slow since Monday,” no numbers. Or worse: no one complains because no one uses it.
• Difficulty: 1/5 · Effort: 1/5.

🟢 Level 2 — Measured#

• Signs: Performance Analyzer used, model size known, unused columns removed (size reduction often 20 to 50% in a model never cleaned up — order of magnitude).
• Gain: we stop guessing. Limit: each measurement is point-in-time.
• Difficulty: 2/5 · Effort: 2/5.

🟡 Level 3 — Managed#

• Signs: we can read the difference between formula engine time (Formula Engine) and storage engine time (Storage Engine); we handle cardinality first (a per-second timestamp over 100M rows ruins compression), then DAX, then the page. Aggregations cover frequent queries.

-- Dans DAX Studio : lire les timings serveur
EVALUATE
SUMMARIZECOLUMNS (
    'DIM_DATE'[Mois],
    'DIM_COMMERCIAL'[Region],
    "Visites", [Nb Visites]
)

• Gain: optimization is targeted. Limit: without a budget, performance degrades version after version.
• Difficulty: 3/5 · Effort: 3/5.

🟡 Level 4 — Budgeted#

• Signs: every release runs a set of reference queries (e.g., 20 DAX queries extracted from Performance Analyzer, replayed by script) and compares durations with the previous version. A threshold breach blocks or alerts.
• Gain: regression is visible before the user sees it.
• Limit: the budget becomes outdated if usage changes.
• Difficulty: 4/5 · Effort: 4/5.

🔴 Level 5 — Guaranteed#

• Signs: concurrent load testing, growth projection, P95 target maintained. Architecture choices are owned: Import for hot data, DirectQuery/Direct Lake for very large volumes, automatic or manual aggregations, partitions.
• Gain: performance stops being a surprise.
• Limit: cost (capacity, CST); freshness/speed trade-off.
• Difficulty: 5/5 · Effort: 5/5.

6. Partial achievement and rating#

Common rule. A “partial 5”: P95 is met under nominal conditions but no load test has been run (BI-PRF-5.1 = 0), growth projection exists (5.2 = 1), P95 target is met on 3 pages out of 5 (5.3 = 0.5); level 5 is not achieved as long as one milestone is at 0.

7. Moving to the next level#

Possible means (optional):

Pitfalls:

• Optimizing the measure that costs 40 ms while the page launches 25 visuals: always look first for the largest contributor.
• Goodhart: a “< 3 s” budget met by hiding slow visuals behind bookmarks: slowness is hidden, not solved.
• Measurements on shared capacity vary with surrounding load: measure in alternation, multiple times, never with a single run.

8. See also#

• Related themes: MOD · DAX · ETL · OPS · CST
• Other matrices: DL — Performance and Scalability
• Content: OneLake + SaaS Mobile Apps: Architecture · SQL Server in SaaS — Challenges and Pitfalls · Comparison — OneLake/Fabric vs Managed SQL Server for Power BI Reporting · Cross-Cutting Power BI Best Practices

✻ Written by Claude Sonnet 5.5