Przejdź do głównej zawartości

Gdy wolumen zmienia wszystko

Dostępne także w 19 językach

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

A wheelbarrow buried under a teetering heap of bricks, its tire flattened, in front of a forklift holding a full pallet of the same bricks with ease.
Obraz wygenerowany przez AI

Moduł 1 z 4 w rozdziale Projektowanie pod skalę, rozdziale 7 kursu Data Architect, i początek części 2. Każdy moduł jest samodzielny. Dalej: Gorące punkty: nierównomierny rozkład własności, rekordów nadrzędnych i wyszukiwań

Pytanie, na które odpowiada ten moduł: co się psuje, gdy dane rosną, i kiedy?

A design can be perfectly reasonable at 70,000 records and completely wrong at 70 million. That’s one of the things that makes data architecture difficult. The object is the same. The fields are the same. The relationships are the same. The business requirement may even be the same. But add a few zeros and the architecture changes. ⟲

Tu przechodzimy od patrzenia do projektowania. W części 1 uczyliśmy się widzieć biznes, model danych, krajobraz systemów przedsiębiorstwa, ład danych i autorytet. Teraz dokładamy tej architekturze ciężaru. Moje pierwsze pytanie brzmi: jak bardzo to urośnie?

Salesforce runs some of the largest CRM implementations in the world, and its core database handles “upwards of 1.1 trillion transactions per month.” Salesforce’s own architects describe that database as “optimized for high-performance transactional workloads on structured data” (The Salesforce Platform: Transformed for Tomorrow (otwiera się w nowej karcie)). It’s built for that: transactional work on structured data. And every record comes with a lot of built-in capabilities and functionality that are part of the core Salesforce platform: the object model, sharing and security, pages, reports, automation, search. All of it runs on a multitenant database, where Salesforce writes the SQL, not you. ⟲

I once taught this class at Dreamforce to a room with a lot of database administrators from Oracle and SQL Server, and of course they were asking, “Why doesn’t Salesforce act this way? Why doesn’t Salesforce do this?” By the end they were smiling, because they’d understood why. Chapter 3 described Salesforce as a multitenant, metadata-driven platform: the apartment building. We’re not tuning our own dedicated database. Salesforce’s storage, indexing, sharing, locking, query optimization and resource protections are designed for many tenants at once. That’s not bad; it’s why we get so much capability without running database infrastructure. But it means the architect needs to understand the platform’s shape. ⟲

Salesforce says it plainly in its own guide: as data sets get larger, “the time required for certain operations grows,” and the architect’s choices change that time “by several orders of magnitude” (Best Practices for Deployments with Large Data Volumes (otwiera się w nowej karcie)). That’s not a weakness. It’s the design. At small volumes you can sometimes ignore the platform’s shape. At large volumes, it reminds you it’s there, and we’ll see that very clearly with selectivity and indexes in the next chapter. That’s why the architect has to know how big this gets. ⟲

The same guide ends with the other half of the story. Salesforce is “a robust environment in which native and custom applications can scale to large volumes of data very quickly while continuing to perform well,” and it names the two ways you get there: make queries selective, and reduce the amount of active data. Millions of records can perform well on Salesforce. They perform well because somebody designed for them. That’s what this part of the course is about. ⟲

When I’m teaching large data volumes, I use a rough ruler. It’s not a Salesforce limit. The numbers are general figures discussed in the community, and Salesforce itself calls a large data volume “an imprecise, elastic term”: tens of thousands of users, tens of millions of records, or hundreds of gigabytes of storage (Best Practices for Deployments with Large Data Volumes (otwiera się w nowej karcie)). And when I say records, I mean records in a single object, not across all your objects. ⟲

Like every number in this course, the ruler has a date on it. More on that in a moment. ⟲

At about 1 million records in an object, there’s just a little bit of drag as dashboards and reports show up. Not a lot, depending on how you filter. Around 10 million, there’s some effect, but it’s still performing. Heading toward 50 million, you start to notice it: list views taking two, three, four seconds, reports and dashboards taking time. Past 50 million, I like to say you’re driving a convertible with your head out the window, and your hair is blown back. Don’t stick your head out the window: past 50 million you’ve got to be streamlined, and you’ve got to know what you’re doing. And once we’re talking about 100 million records or more, I start asking a different question: should these records be here at all? I’ve heard of people running a hundred million or more in an object. But if you’re at 50 million today, you’re going to 100, 150, 200 million in the years ahead. If you’re not archiving today, when will you? If somebody tells me they have 150 million records in an object and almost nobody ever looks at the old ones, that’s not an operational object any more. That’s an audit log. ⟲

To nie przypadek, że miarka zaczyna się od miliona. Jak pokaże następny rozdział, własne wyliczenia zapytań platformy też zmieniają się przy „pierwszym milionie rekordów”. Ale dokładne liczby nie są tu najważniejsze. Najważniejsze jest to: wolumen zmienia projekt.

The LDV ruler The LDV ruler, four steps in records in a single object. One million: a little drag; dashboards and reports take a moment, depending on how you filter. Ten million: some effect, but still performing. Fifty million: the convertible; list views take two, three, four seconds, reports and dashboards take time, and past it your hair is blown back. One hundred million and up: should it be here? When will you archive, and is this operational data or an audit log? Underneath: records in a single object; a rule of thumb from the community, not a Salesforce limit. {"author": "Steve Simpson", "blueprint": "1.1.0", "builder": "standard 1.53.0", "built": "2026-10-29", "id": "DGM-DA-LDV-11", "rights": "\u00a9 2026 Twinshark LLC. All rights reserved.", "template": "comparison 0.5.0", "theme": "stevetecharc 1.4.0"} 1 million a little drag YOU NOTICE Dashboards and reports take a moment Depends on how you filter 10 million some effect YOU NOTICE Some effect Still performing 50 million the convertible YOU NOTICE List views take 2, 3, 4 seconds Reports and dashboards take time Past it, your hair is blown back 100 million + should it be here? YOU ASK When will you archive? Operational data, or an audit log? Records in a single object. A rule of thumb from the community, not a Salesforce limit. © 2026 Twinshark LLC · stevetecharc.com
Kliknij, aby powiększyć

Every zero changes what you notice. A rule of thumb, not a limit. ⟲

Salesforce ships three releases a year. Over time the platform grows, some limits are raised, and systems get faster. So I don’t want you to treat any number in this course as if it were permanent, including mine. Salesforce keeps every release’s edition of its guides online, which means we can do something useful: put ten years side by side. Here’s Winter ‘16, from late 2015, next to Winter ‘27, from 2026. ⟲

Winter ‘16 (2015)Winter ‘27 (2026)
Big objectsDidn’t exist yet“Consistent performance, whether you have 1 million records, 100 million, or even 1 billion”
Skinny table columns100200
Index selectivity thresholdsCapped at 1 million records (standard) and 333,333 (custom)No published cap, and the optimizer can add an index on its own
Minimum data storage1 GB per org, or 20 MB per user10 GB per org, plus storage per user
Bulk API batches a day5,00015,000
Parent-to-child levels in one SOQL query15
Apex heap per transaction6 MB (12 MB asynchronous)10 MB (25 MB asynchronous)

That’s the upward direction: limits raised, caps removed, new kinds of storage, and an optimizer that now “recognizes that an index can improve performance for frequently run queries” and adds one (Working with Very Large SOQL Queries (otwiera się w nowej karcie)). ⟲

Now look at what didn’t move in those ten years. The definition of a large data volume is word for word the same: “tens of thousands of users, tens of millions of records, or hundreds of gigabytes.” The advice to avoid having any user own more than 10,000 records is the same. And an Apex transaction still gets 100 SOQL queries, 50,000 records and 10 seconds of CPU (Execution Governors and Limits (otwiera się w nowej karcie)). Those are the house rules of the apartment building, the guardrails that keep one tenant from hurting the others, and they’ve barely moved. ⟲

Salesforce is open about the pace, too. It ships three releases a year, each with a new API version, and it commits to “supporting each API version for a minimum of 3 years from the date of first release,” with at least a year’s notice before support for a version ends (SOAP API End-of-Life Policy (otwiera się w nowej karcie)). In practice versions live much longer: as I write this, version 41.0, from Winter ‘18, is still supported. If you’re new, you don’t need the history. But it does change decisions. The org you inherit was designed under older limits, and some of its workarounds may not be needed any more. The integration you build today is pinned to an API version with a clock on it. And study material written a few releases ago may still teach yesterday’s numbers. ⟲

So when you read a number in this course, read the date next to it. The specific limits will keep changing, mostly upward, and Salesforce will keep pivoting. The thinking won’t: know your Number of Zeros, keep the active data small, make your queries selective, and check the current edition of the guide before you commit to a number. That’s why I teach the ruler as a ruler, not a law. For now, it’s a good place to start. ⟲

The sources, both editions: Best Practices for Deployments with Large Data Volumes, Winter ‘16 (otwiera się w nowej karcie) (archived copy (otwiera się w nowej karcie)) and Winter ‘27 (otwiera się w nowej karcie) (archived copy (otwiera się w nowej karcie)); Salesforce Limits Quick Reference Guide, Winter ‘16 (otwiera się w nowej karcie) (archived copy (otwiera się w nowej karcie)), and Salesforce Developer Limits and Allocations Quick Reference, Winter ‘27 (otwiera się w nowej karcie) (archived copy (otwiera się w nowej karcie)); Data and File Storage Allocations (otwiera się w nowej karcie) (10 GB “starting in late March 2019”); Execution Governors and Limits (otwiera się w nowej karcie), Winter ‘27. ⟲

One of the easiest mistakes is sizing an architecture from today’s data. Steve Trucking has about 1 million deliveries a year. If I only ask “can Salesforce handle one million deliveries?”, I’m asking the wrong question. Steve Trucking expects about 30% annual growth: 1.0 million in year one, 1.3 in year two, 1.69 in year three, about 2.2 in year four, about 2.86 in year five. That’s annual volume. If we keep all five years, we’re not designing for 2.86 million deliveries; we’re designing for roughly 9.04 million. And that’s before we ask how much history already exists. ⟲

Dlatego uczę Liczby Zer Steve’a (Steve’s Number of Zeros). Nie zaczynaj od fałszywej precyzji. Znajdź właściwy rząd wielkości. Dziesięć tysięcy, sto tysięcy, milion, dziesięć milionów, sto milionów? Każde dodatkowe zero może zmienić architekturę.

The delivery isn’t the only record. Steve Trucking averages about 1.6 line items a delivery, so over five years that’s about 14.5 million work order line items. And there are around eight status changes a delivery: about 72.3 million status events. The requirement sounded like “track deliveries.” The architecture has to deal with tens of millions of related records. One million parents with eight children each is not a one-million-record problem. You have to follow the model. ⟲

Checkpoint 1 also gave us four to five million photos a year, and FleetLink produces about 200,000 GPS readings a day: roughly 73 million a year before growth. That’s why we decided not to put every GPS reading, every photo or every status change into the Salesforce operational model. We hadn’t even started this chapter, and the Number of Zeros was already telling us something. It’s like telling the doctor, “It hurts when I raise my arm,” and the doctor says, “Then don’t raise your arm.” Sometimes the best way to solve a large-data-volume problem is don’t create it in the first place. ⟲

When people hear large data volumes, they think “queries get slow.” Yes, and the next chapter is all about query and retrieval. But it’s bigger than that. As volume grows, I care about sharing calculations, ownership, record locking, parent-child relationships, list views, related lists, reports, search results, batch processing, integrations, data loads, deletes, archiving, backups and recovery. Even a harmless-looking decision gets expensive when it’s repeated millions of times. ⟲

That’s the lesson: what works today may fail tomorrow. Maybe nobody notices at 50,000 records, or at 500,000. Then growth pushes the object through several million, and an old design assumption becomes the bottleneck. The system didn’t randomly go bad. The architecture crossed the scale where that decision mattered. ⟲

My starting position is still: design clean first, then back off when necessary. Understand the entities and relationships; don’t duplicate data because it’s convenient. But normalization is a principle, not a rule. If a report or operational query constantly has to walk several relationships across tens of millions of records to get one value, at some scale I may copy that value closer to where it’s used. That’s denormalization. ⟲

Załóżmy, że raportowanie pozycji zleceń pracy według partnera zawsze potrzebuje partnera detalicznego (Retail Partner), a w czystym modelu jest on dwa skoki dalej: pozycja → zlecenie pracy → partner. Przy wystarczającym wolumenie może być powód, żeby umieścić zarządzaną kopię identyfikatora partnera bezpośrednio na pozycji. Teraz stworzyłem duplikację i razem z nią biorę na siebie odpowiedzialność. Jak jest wypełniana? Czy może się zmienić? Jak pozostaje dokładna? Która kopia jest nadrzędna? To realne koszty, a ja czasem akceptuję kontrolowaną duplikację, żeby uzyskać wydajność, której potrzebuje biznes. Projektuj w stronę normalizacji, a potem świadomie się wycofuj, gdy wolumen da ci powód. Nie denormalizuj, bo tak łatwiej zbudować; denormalizuj, bo potrafisz wyjaśnić korzyść i koszt, który akceptujesz.

Widzę jeszcze jeden błąd związany z wolumenem. Biznes chce raportów, więc do operacyjnego CRM dokłada się historię, i jeszcze, i jeszcze, aż uzasadnieniem staje się „do raportowania”. Wtedy pytam: czy zamieniamy system operacyjny w system analityczny? To dwa różne obciążenia. Steve Trucking potrzebuje Salesforce do prowadzenia dzisiejszych dostaw: obsługa klienta potrzebuje dzisiejszego statusu, planowanie dzisiejszych tras, partnerzy dzisiejszych postępów. Analityk porównujący pięć lat wyników dostaw według regionu, partnera, ciężarówki, pogody, ekipy i okna dostawy, i zestawiający to z danymi finansowymi, prowadzi obciążenie analityczne. Salesforce potrafi raportować dane operacyjne. To nie znaczy, że każde historyczne pytanie analityczne należy do magazynu operacyjnego. Steve Trucking ma już Snowflake. Zapytaj więc: na które pytania musi odpowiadać system operacyjny, a które należą do platformy analitycznej? Część 3 omawia hurtownie danych dokładnie.

Migawki raportowe to też nie strategia hurtowni danych. Spłaszczone obiekty raportowe, rekordy migawek, wartości powielone dla wygody raportowania: czasem są uzasadnione. Mogą też być sposobem na odłożenie pytania architektonicznego. Jeśli tworzę kolejne miliony rekordów tylko po to, żeby raportować miliony rekordów, które już istnieją, chcę wiedzieć, gdzie to się kończy. Nie produkuj dalej rekordów operacyjnych tylko dlatego, że nikt nie chce rozmawiać o hurtowni danych.

Dwa systemy mogą mieć po 20 milionów rekordów i zachowywać się zupełnie inaczej. Jeden to 20 milionów rekordów historycznych, prawie nigdy nieaktualizowanych, zwykle odczytywanych pojedynczo po znanym identyfikatorze, z prostym udostępnianiem. Drugi to 20 milionów aktywnych rekordów, stale aktualizowanych, ze złożoną własnością, mnóstwem raportów, wieloma użytkownikami, silnym skupieniem nadrzędny-podrzędny i częstymi integracjami. To nie jest ten sam problem. Dlatego sama liczba rekordów nie wystarcza. Jak szybko rośnie? Jak często się zmienia? Jak jest odpytywany, udostępniany, powiązany? Jak długo pozostaje aktywny? Ile napływa lub zmienia się naraz? Co się dzieje przy usuwaniu? „Salesforce poradzi sobie z 20 milionami rekordów” to nie odpowiedź. Ciekawe pytanie brzmi: 20 milionów rekordów, które robią co?

Agents can create data very quickly: interactions, tasks, summaries, recommendations, logs, intermediate results, events, audit information. Maybe that’s useful. But AI doesn’t make storage free, or sharing calculations, or indexing, and it certainly doesn’t make useless history valuable. If an agent creates ten records where a human process used to create one, our volume assumptions may be wrong before the business itself grows. I’m not yet seeing agents generate more records in the orgs I work with, but it’s the natural progression of more and more data being tracked. ⟲

Poza egzaminem: co zmienia AI. AI może znacznie zwiększyć tempo, w jakim systemy tworzą i pobierają dane. Modeluj jawnie wolumen generowany przez agentów, zamiast zakładać, że historyczne tempo transakcji ludzi przewiduje przyszły wolumen.

At the end of Part 1 we made several decisions that are now reinforcing each other. The work order is central. The current status lives in the org, and the history of every status change is published to the warehouse, Snowflake. Photos live outside the core. Raw GPS stays with FleetLink. At five years of 30% growth, we’re approaching 9 million work orders, 14.5 million line items and 72.3 million status events, and hundreds of millions of GPS readings if we’d copied them all. ⟲

Was keeping raw GPS out of the operational model just an integration decision? No; it was a scale decision too. Was putting photos off-core just a storage decision? No; scale again. And the status history? It has real business value, so we keep it. But at 72.3 million events, keeping it outside the operational database is the difference between an org that carries its history and an org that’s buried by it. The question that remains for what we do keep in core: how long should those records stay active? That’s module 4. ⟲

Steve Trucking's zeros at year five Steve Trucking's zeros at year five. In Salesforce: 9.04 million work orders, from 1 million a year growing 30%, with five years kept; and 14.5 million line items, at 1.6 per work order. Off core: 72.3 million status events, at 8 per work order, in Snowflake, while the org keeps the current status; and 73 million GPS readings a year, 200,000 a day before growth, in FleetLink, while the org keeps the latest position. Underneath: work orders a year are 1.0, 1.3, 1.69, 2.2 and 2.86 million; the biggest zeros were kept out of core on purpose. {"author": "Steve Simpson", "blueprint": "1.0.0", "builder": "standard 1.28.0", "built": "2026-10-15", "id": "DGM-DA-LDV-12", "rights": "\u00a9 2026 Twinshark LLC. All rights reserved.", "template": "comparison 0.3.0", "theme": "stevetecharc 1.4.0"} 9.04 million work orders HOW 1 million a year, growing 30% Five years kept LIVES In Salesforce 14.5 million line items HOW 1.6 per work order LIVES In Salesforce 72.3 million status events HOW 8 per work order LIVES In Snowflake; the org keeps the current status 73 million a year GPS readings HOW 200,000 a day, before growth LIVES In FleetLink; the org keeps the latest position Work orders a year: 1.0, 1.3, 1.69, 2.2, 2.86 million. The biggest zeros were kept out of core on purpose. © 2026 Twinshark LLC · stevetecharc.com
Kliknij, aby powiększyć

Follow the model, and the zeros multiply. ⟲

  • Jak dziś wygląda Liczba Zer, łącznie z oczekiwanym wzrostem i całym okresem retencji, a nie tylko czy dzisiejszy wolumen działa.
  • Które czyste struktury pozostają znormalizowane, a gdzie wolumen uzasadnia kontrolowaną denormalizację.
  • Które dane zostają w operacyjnym core Salesforce, a które należą do wyspecjalizowanego magazynu, platformy analitycznej lub innego systemu.
  • Czy wymaganie raportowe jest naprawdę operacyjne, czy to sygnał, że biznes potrzebuje architektury analitycznej.
  • Jak nowi producenci danych, tacy jak agenci, zmieniają założenia co do wolumenu, zanim ich rekordy staną się jutrzejszym problemem.

Large data volume architecture isn’t about being afraid of big numbers. Salesforce runs very large implementations. One million records isn’t automatically a problem; neither is ten million or a hundred million. But every zero should make me ask another question: how will we find them, who owns them, how are they shared, what are they related to, how often do they change, how long do we need them, and where should they live? In Part 1 we learned to recognize the right entities and relationships. Now we make sure those bones can carry the weight, because what works today may fail tomorrow. ⟲