When Salesforce Uses an Index

🌐 Also in 19 languages

Español · 日本語 · Português (Brasil) · Français · Deutsch · Italiano · 한국어 · 简体中文 · Português (Portugal) · 繁體中文 · Bahasa Indonesia · Türkçe · Polski · Tiếng Việt · ไทย · Українська · Español (México) · Français (Canada) · Deutsch (Schweiz)

Image generated with AI

Module 3 of 5 in the chapter Query and Retrieval Architecture, chapter 8 of the Data Architect course. Each module stands on its own. Previous: How Salesforce Finds Your Data · Next: Indexes, Skinny Tables and Redesign

The question this module answers: When can Salesforce use an index?

In the last module we built the card catalog: what an index is, the cost of every copy on every save, and which fields Salesforce indexes for you, which ones you can ask for, and which ones can’t have an index at all. So now we have an index. Can it be used in all circumstances? No. Let me tell you a story first.

The Nitro Boost

Picture a street race. Two cars side by side, engines roaring. Near the end, one driver presses a button and the car jumps forward and wins. That button is the nitro: a boost of extra fuel that gives the engine a burst of speed for a few seconds.

But the nitro doesn’t fire just because you press the button. A nitro system is usually built so it only fires when three things are true at once:

1. The system is installed and switched on. No nitro in the car, no boost.
2. The driver has the pedal all the way down. Press it halfway, and the system stays off.
3. The engine is in the right speed range. Too slow or too fast, and the system won’t fire.

All three, and the car jumps ahead. Miss any one, and you’re just driving.

An index is your query’s nitro. And it works the same way: three things have to be true before Salesforce uses it.

The car	Your query
The nitro is installed and switched on	There’s an index on the field you filter on
The pedal is all the way down	You ask in a selective way: equals, starts with
The engine is in the right range	Not too many records match: you’re under the threshold

This module takes them one at a time. Let’s start with the first.

Is There an Index?

Let’s start with the question the catalog was built for. Walk up to the desk and ask for Plumbing Basics. The librarian goes to the drawer of cards by title, finds the card, reads the shelf (B4), walks over and takes the book down. One card, one book. In a library of 10 million books, that’s one card and one book instead of 10 million books. That’s what an index is for, and it’s what Salesforce does when you filter on an indexed field: find the entry, read which record it points to, go straight there.

Now ask a different question: “Can I have every book from Harbor Press?” We have drawers by title, by subject and by author. Is there a drawer by publisher? No. So what does the librarian do? Walks every aisle and opens every book. That’s a table scan, and the three drawers we paid for didn’t help at all.

In Salesforce it’s the same. Filter on a field that isn’t indexed, and the indexes on the other fields can’t help that filter. That’s why the last module’s tables matter: before you worry about anything else, check whether the field you filter on has an index, standard or custom. That’s the first of the three: is the nitro installed?

Click to enlarge

By title, there’s a drawer: find the card, get the book. Fast.

Click to enlarge

By publisher, there’s no drawer: no match in the index. Go around it and open every book: a table scan. Slow.

Asking the Index the Right Way

Now say there is a drawer. The title drawer, sorted A to Z. Which questions can it answer quickly?

• Find the book titled Plumbing Basics. Go to P, find the card. Fast.
• Find every book whose title starts with “Plumbing”. Go to P, read the cards that start with “Plumbing”, stop when they don’t. Fast.
• Find these three titles. Three trips to the drawer. Fast.
• And with a drawer of cards by date, find every book added this year: go to the first card for this year and read until the year ends. Fast.

Each of those questions tells the librarian where to start in the drawer. Now some questions the same drawer can’t answer that way:

• Find every book whose title contains “Plumbing”. “Advanced Plumbing” is filed under A. “Home Plumbing Guide” is filed under H. The drawer is sorted by how a title starts, so there’s no place in it to start looking. The drawer can’t answer, and the librarian is back to walking every shelf.
• Find every book whose title ends with “Basics”. Same problem: the end of a title tells you nothing about where it’s filed.
• Find every book that isn’t Plumbing Basics. Nearly every card matches. The drawer narrows nothing.
• Find every book with no title. In Salesforce’s drawers, there’s no card for an empty value unless someone asked for one (we saw that in the last module).

None of these tells the librarian where to start. In class I say it this way: indexes work from left to right. Ask for the start of the value and the index can help. Ask for the middle, the end, or everything except one value, and it can’t: no match in the index, and Salesforce falls back to a table scan.

Click to enlarge

Starts with: straight to P. One card, one book. Fast.

Click to enlarge

Contains: no way into the drawer, no match in the index. Go around it and open every book: a table scan. Slow.

Here’s the same list in SOQL:

Ask it this way	In SOQL	Can the index help?
Equals	Name = 'Plumbing Basics'	Yes
Starts with	Name LIKE 'Plumbing%'	Yes
One of a short list	Id IN ('a01…', 'a02…', 'a03…')	Yes
A range of dates or numbers	CreatedDate = THIS_YEAR, Amount > 10000	Yes
Contains	Name LIKE '%Plumbing%'	No
Ends with	Name LIKE '%Basics'	No
Not equal, not like, not in	Name != 'Plumbing Basics'	No
Empty	Region__c = null, Name != ''	No, by default

That’s not just my rule of thumb. Salesforce lists the cases where a custom index typically isn’t used, including “The filter operator is a negative operator such as NOT EQUAL TO (or !=), NOT CONTAINS, and NOT STARTS WITH” and “You’re comparing with an empty value (Name != ”)” (Working with Very Large SOQL Queries (opens in a new tab), Apex Developer Guide, Winter ‘27). And: “Custom indexes are never used for the following operations: NOT EQUAL TO comparisons, null value comparisons (e.g., Name = ”), leading wildcard searches (e.g., Name LIKE ‘%Smith’)” (Developer Console Query Plan Tool FAQ (opens in a new tab), Help, version 5).

And if what you really need is to find a word anywhere in the text, remember the two ways to ask from the last module: that’s what FIND and the search engine are built for. Module 5 covers it.

The Word for It: Selective

So there are two things to check before an index can speed up your query. One: is there an index on the field? Two: are you asking for the data in a way the index can use? Equals, starts with, a short list, a range of dates or numbers: yes. Contains, ends with, not equal, empty: no.

When you ask the second way, I say you’re asking in a selective manner. That’s the word to learn: selective. A selective filter is one Salesforce can answer through an index instead of walking every record. Reports, list views and SOQL all care about it.

Salesforce’s own definition has one more part, and it’s the reason this module isn’t over yet: “A query is selective when one of the query filters is on an indexed field and the query filter reduces the resulting number of rows below a system-defined threshold” (Working with Very Large SOQL Queries (opens in a new tab), Winter ‘27). An index, asked the right way, and not too many records. That’s the second of the three, the pedal all the way down. The third check, the engine in the right range, is where people get into trouble, and it comes next.

Test knowledge. A filter can use an index only if the field is indexed and you ask in a selective way. Selective: equals, starts with (LIKE 'abc%'), a short IN list, a range on a date or number. Not selective: contains and ends with (a leading %), negative operators (!=, NOT LIKE, NOT IN), and comparisons with an empty value.

A note on encrypted fields. If a field is encrypted with Shield Platform Encryption, check before you build an important filter on it. Salesforce’s own Well-Architected pages don’t agree today. Trust (opens in a new tab) says probabilistic encryption can’t be used to filter, sort or aggregate, and that deterministic encryption allows exact-match filters. Data Protection Patterns (opens in a new tab) says probabilistic encryption “supports all query operations”. Read both (checked 2026-10-21) and test the filter in your own org.

Everybody Lives in Arizona: When an Index Won’t Help

So you have an index, and you’re asking it the right way: equals, nothing fancy. Is that enough? Not always. Here’s the example I use in class.

Say we have the addresses of 10 million people from all over the United States, and the state field is indexed. Ask for everyone in Arizona: State = 'AZ'. Arizona is a small slice of the whole country, so the index narrows 10 million records down to a small group. Go to the drawer, pull the Arizona cards, get those records. Fast. That’s the index doing its job.

Now change the data, not the query. Same field, same index, same State = 'AZ'. But this time everybody lives in Arizona. Every one of the 10 million records matches. The index doesn’t narrow anything. Going to the index first, then fetching 10 million records one card at a time, is actually slower than just reading the whole table. The index isn’t broken. It just isn’t useful for this question.

And here’s the key: it depends on what you’re looking for. Keep the all-Arizona data and add one person in Texas. Ask for State = 'TX': one match. The index takes you straight to that one record. Fast. Ask for State = 'AZ' on the same data: almost everything matches. Slow. Same field, same index, same data. The only difference is how many records match what you asked for. When a large number of records match what you’re looking for, the index stops helping.

Watch for this with fields like case status. Status = 'Closed' can be a good filter when you launch, and stop being one as closed cases pile up year after year. We’ll come back to that in The Closed-Cases Trap.

Click to enlarge

People from all over the US: Arizona is a small slice. Fast.

Click to enlarge

Everybody lives in Arizona: every card matches. Reading the table is faster. Slow.

Click to enlarge

The same data, one person in Texas: one card. Fast. What you’re looking for decides.

Why Too Many Matches Is Slower

Think about the library again. You want one book: find the card, walk to the shelf, done. Now you want 9 million of the 10 million books, and you insist on using the card catalog. Find a card, walk to the shelf, come back to the drawer, next card, walk to another shelf, come back… nine million times. It would be faster to start at the first aisle and walk the shelves once, in order.

Salesforce builds that limit into the index itself: “The index table places upper limits on the number of records that an indexed search can effectively return” (Indexes (opens in a new tab), Winter ‘27). So before using an index, Salesforce counts: how many records would this filter match? If the answer is a small part of the object, the index is worth it. If the matches rise above a certain percentage of the records, Salesforce skips the index and reads the table.

That percentage has a name: the selectivity threshold. Remember Salesforce’s definition of a selective query: an indexed field, asked the right way, and a filter that “reduces the resulting number of rows below a system-defined threshold”. This is that threshold, and it’s the third check. Let’s look at the numbers.

The Thresholds

Why does the threshold matter so much? Salesforce’s Well-Architected guidance doesn’t soften it: “Query selectivity is the single largest lever on whether a solution scales to millions of records or times out at hundreds of thousands.” And: “Selectivity is capacity you don’t have to buy” (Resource and Cost Optimization (opens in a new tab)). An efficient query leaves room for every other tenant in the building.

Salesforce sets the line as a percentage of the records, and the percentage depends on the kind of index: one rule for a standard index, a lower one for a custom index (Make Salesforce Platform SOQL Query Selective (opens in a new tab), Help, version 4). Let’s start with the simple case and build up.

Start Simple: a Standard Index

Take the Name field. It has a standard index. Say we have 100,000 accounts. For a standard index, the line is 30% of the records: 30,000.

Now ask for Name LIKE 'Steve%'. Steve’s Plumbing, Steve’s Garage, a few others: say 2,000 accounts. 2,000 is under 30,000, so the filter is selective and Salesforce uses the index. Fast.

Same field, same index, same 100,000 accounts. Now ask for Name LIKE 'S%', every account that starts with S. Say that’s 40,000. 40,000 is over the line. Not selective: Salesforce skips the index and reads the table.

Click to enlarge

Same index, same records. Under the line, the index is used. Over it, it isn’t.

Both filters ask in a selective way, a “starts with”. Both use an indexed field. The only difference is how many records match.

As the Object Grows

So the line is 30%? For the first million records, yes. After that, Salesforce counts only 15% of the records beyond the first million, and the line never goes above 1 million. In Salesforce’s words, the standard threshold “maxes out at 1 million total targeted records, which you could reach only if you had more than 5.6 million total records” (same source).

• 500,000 records: 30% of them, 150,000.
• 1 million records: 30% of them, 300,000.
• 2 million records: 300,000 for the first million, plus 15% of the second million (150,000): 450,000.
• 6 million records: the math would give 1,050,000, but the cap is 1 million. That’s the top end.

Click to enlarge

The line grows with the object, more slowly after the first million, and stops at 1 million.

Look at what happens to the share. At 1 million records, a filter can match 30% and still be selective. At 2 million, 22.5%. At 6 million, under 17%. The bigger the object, the more a filter has to narrow it down. Salesforce’s Well-Architected guidance puts it the same way: the thresholds “tighten as an object grows past its first million records” (Resource and Cost Optimization (opens in a new tab)).

Custom Indexes: a Lower Line

Now a custom index. Back to State = 'AZ': State isn’t a standard index, so it needs a custom one. A custom index gets a lower line: 10% of the first million records, plus 5% of the records beyond it, never more than 333,333 (same source).

• 100,000 records: 10% of them, 10,000.
• 1 million records: 10% of them, 100,000.
• 6 million records: the math would give 350,000, but the cap is 333,333. The top end.

Click to enlarge

A custom index: a third of the standard line, and a cap of 333,333.

So on a million contacts, State = 'AZ' is selective only if fewer than 100,000 of them live in Arizona. And on 10 million contacts, the line is still 333,333, about 3.3%. That’s the part people forget: eventually the thresholds stop growing. That’s why large-volume design isn’t “put an index on it.” The question is what will the distribution look like at scale?

Test knowledge. Know the two thresholds. Standard index: 30% of the first million records, plus 15% of the rest, up to 1 million. Custom index: 10% of the first million, plus 5% of the rest, up to 333,333. Be ready to work one out: a standard index on 2 million records allows 450,000; a custom index on 1 million allows 100,000.

The Closed-Cases Trap

One of my favorite examples, because the query works perfectly at launch. A filter on IsClosed = true. Early on: 100,000 cases, 20,000 closed. Great. Time passes, and cases have an interesting habit: they close. Now 10 million cases, 9 million closed. Nothing about the query, the field or the code changed. The distribution did. The filter that once narrowed the population now describes almost all of it. It crossed from selective to non-selective because the data aged. A query can become bad without the query changing.

Add 30% Growth

So I combine selectivity with the Number of Zeros and growth. If the customer has no projection, I use 30% a year as a planning baseline, not because every business grows 30%, but because it gives us something to design against instead of pretending today’s volume is permanent. Start at 1.5 million records: 1.95 million in year two, 2.54 in year three, 3.30 in year four, 4.28 in year five, 5.57 in year six. At about 5.6 million records, the uncapped custom threshold is 100,000 plus 5% of 4.6 million, 330,000: essentially the 333,333 cap. From there the data keeps growing and the custom threshold effectively doesn’t. Do that math before production teaches it to you.

Click to enlarge

Today’s selective filter can become tomorrow’s production problem.

Two different 30%s. Don’t mix up this 30% with another 30% you’ll meet in Salesforce’s Well-Architected guidance. Reliability (opens in a new tab) says: “Build margin by targeting 70% of governor limits as the operational ceiling under normal conditions, reserving 30% for unexpected spikes.” That 30% is room to spare under the governor limits of a transaction. The 30% here is how fast the data grows each year. Different numbers for different things. And notice how Salesforce describes the 70% figure on another page: “a widely cited community best practice, not an official Salesforce specification” (The Agentic Enterprise (opens in a new tab)). Treat it as good advice, not a rule (pages checked 2026-10-22). And remember: when your main records grow by 30%, the records under them can grow faster. Each new account brings orders, each order brings line items, and every one of them brings activities and history. Child objects and transaction data can multiply the growth, and you can’t always tell which one will grow toward a Salesforce limit first, or past it. So do the math for the children, not only the parents.

Steve Trucking: Today’s Query, Five Years On

Launch with a million deliveries; coordinators often filter on Status = 'Scheduled', about 100,000 records, 10%. Reasonable. Five years later there are about 9 million cumulative work orders, but the number scheduled at any moment is driven by the operating window, so it stays small. That filter may actually get more selective as history grows. Now Status = 'Completed': completed records pile up year after year, and that filter gets steadily worse at narrowing anything. Same field, same index, different value, different distribution, different behavior. That’s selectivity.

Beyond the exam: what AI changes. An AI assistant writes WHERE Status = 'Completed' as easily as a developer does, and it works in a test org. Production still has the same optimizer, indexes, thresholds and distribution. AI doesn’t repeal database physics. Test AI-generated queries against realistic volume and distribution, not 500 sample cases.

All Three: the Boost

Back to the race. Three things have to be true for the nitro to fire, and now you know all three for a query:

1. There’s an index on the field you filter on. (The nitro is installed.)
2. You ask in a selective way: equals, starts with, a short list, a range of dates or numbers. (The pedal is all the way down.)
3. Not too many records match: you’re under the threshold for that index, today and as the data grows. (The engine is in the right range.)

All three, and Salesforce goes to the index and your query jumps ahead. Miss any one, and Salesforce reads the table: you’re just driving. And notice that the third one can change on its own. The closed cases pile up, the object grows, and a query that had the boost last year doesn’t have it this year.

So who checks all three, every time a query runs? Not you, and not the developer. That’s the job of the librarian, and that’s the next module.

Test knowledge. For an index to be used, three things must be true: the field is indexed, the filter is selective in the way it asks (no leading wildcard, no negative operator, no comparison with an empty value), and the number of matching records is under the selectivity threshold.

What an Architect Decides Here

• Which high-volume filters must stay selective today, and at year three, year five and beyond.
• Whether the fields those filters use have an index, standard or custom, and whether one is worth its cost.
• Whether the filters ask in a selective way, or rely on contains, ends with, not equal or empty values.
• Whether the real distribution of values, not just the presence of an index, keeps each filter under its threshold as the data grows.
• Which queries, including AI-generated ones, need testing at realistic volume and distribution before production.

Yes, know the thresholds: standard, 30% of the first million plus 15% beyond, capped at 1 million; custom, 10% plus 5%, capped at 333,333. But if that’s all you learn, you missed the lesson. The card catalog is useful only when it saves enough work; an indexed filter that returns almost everything isn’t a shortcut; and the same filter can drift from selective to non-selective as data grows and ages. So when someone says “don’t worry, that field is indexed,” ask the three questions. Then do the math.

Explore

• Salesforce Developers, Apex Developer Guide (Winter ‘27), Working with Very Large SOQL Queries (opens in a new tab), checked 2026-11-01: what makes a query selective, and when a custom index typically isn’t used.
• Salesforce Help, Make Salesforce Platform SOQL Query Selective (opens in a new tab) (version 4, April 2026: the thresholds and their caps) and Developer Console Query Plan Tool FAQ (opens in a new tab) (version 5, June 2026), checked 2026-11-01.
• Salesforce Architects, Well-Architected, Resource and Cost Optimization (opens in a new tab), checked 2026-11-01: Query Selectivity.
• Salesforce Developers, Best Practices for Deployments with Large Data Volumes (Winter ‘27), Indexes (opens in a new tab), checked 2026-11-01: the index table places upper limits on how many records an indexed search can return.
• Salesforce Architects, Well-Architected, Trust (opens in a new tab) and Data Protection Patterns (opens in a new tab), checked 2026-11-01: filtering on encrypted fields (the two pages differ).
• Salesforce Architects, Well-Architected, Reliability (opens in a new tab) and The Agentic Enterprise (opens in a new tab), checked 2026-11-01: the 70% governor-limit ceiling and its 30% reserve, and how Salesforce describes it.