Category: Big Data

Many companies choose a database the way they choose an electricity supplier — once, quickly, and then never again until something goes dark. The decision gets made by whoever was on the project that week, and everyone else lives with it for the next eight years.

That’s backwards. Your database quietly sets the ceiling on four things at once: how fast your team ships, how honestly you can report, how well you survive an audit, and what it costs you to double in size. None of those are IT problems. They’re business problems that happen to be expressed in SQL.

The good news is that the shortlist in 2026 is genuinely good. The bad news is that “good” isn’t the same as “right for you.” Below is a workload-first way to think about it, followed by eight systems worth knowing — and the honest tradeoff each one asks you to accept.

Only question that matters first

Before you compare vendors, answer this: what does this data actually do?

• It records commitments — orders, payments, contracts, payroll. This data needs to be correct more than it needs to be fast. A transaction either fully happens or fully doesn’t.
• It describes things loosely — product catalogs, customer profiles, logs, anything whose shape changes every quarter because a PM changed their mind.
• It answers questions — dashboards, forecasts, cohort analysis, the report leadership stares at every Monday.
• It needs to be instant — session tokens, live inventory, pricing lookups measured in single-digit milliseconds.

Those four jobs want four different architectures. Most technical debt in data infrastructure comes from a single decision: using one system to do all four because it was already there.

Workload 1: Recording commitments (relational)

This is where ACID compliance stops being jargon and starts being the reason your balance sheet reconciles.

PostgreSQL

If a single database has won the last decade, it’s this one. Postgres delivers strict ACID guarantees, first-class JSONB for the semi-structured 20% of your data, serious indexing, and an extension ecosystem that lets it stretch into time-series, geospatial, or vector search without leaving the box. The cost story is hard to argue with: the engine is free, and every major cloud sells a managed version (RDS, Cloud SQL, Azure Database) so you’re not paying people to babysit failover. Instagram famously runs a sharded Postgres cluster at a scale that should have killed the “SQL doesn’t scale” line permanently. Wise and Revolut keep money in it. That’s a decent reference list. Choose it when: you want capability without lock-in, and you’d rather not re-platform in three years.

MySQL

Still one of the most deployed databases on earth, and the one most small businesses meet first. It’s cheap to host, quick to stand up, and compatible with essentially every framework and CMS ever written. Where it strains is complex analytical queries and heavy concurrent writes. For new enterprise builds, Postgres has largely taken the oxygen. For read-heavy, content-driven applications, MySQL is a perfectly rational choice and will be supported forever.

Microsoft SQL Server

If your company already lives inside Azure, Power BI, .NET, and Active Directory, SQL Server removes an enormous amount of integration friction. Always On Availability Groups give you genuine HA/DR, which matters if downtime is a regulatory event rather than an inconvenience. The catch is licensing, which is not cheap and not getting cheaper. Managed Azure SQL at least absorbs the patching and failover work you’d otherwise staff for.

Oracle

Still the system of record for a large share of the world’s banks, insurers, and government agencies — and not out of nostalgia. Real Application Clusters handle concurrency at a level few competitors reach, and the audit trail and row-level security tooling is mature in a way that only decades of regulated deployments produce. What’s changed is the shape of Oracle estates. The common 2026 pattern isn’t ripping Oracle out; it’s shrinking it. Core ledger stays. The forty surrounding systems that never needed Oracle-grade guarantees move to Postgres, and the licensing bill drops accordingly.

Workload 2: Loose shapes and instant reads (NoSQL)

SQL versus NoSQL was never a war. It’s a question about the shape of your data.

MongoDB

The document model stores records as JSON-like objects, which suits anything where records don’t share a fixed template: catalogs with wildly different attributes, customer profiles that grow fields monthly, nested logistics hierarchies. The practical payoff is migration scripts you don’t have to write. Your team changes the model and moves on. Barclays, Bosch, and Adobe run it at scale — it stopped being a startup toy a long time ago.

Redis and Valkey

This one needs a 2026 update, because the in-memory landscape genuinely changed. Redis moved from BSD to SSPL in March 2024, which ended its status as open source in the OSI sense. The community forked within days: Valkey, now under the Linux Foundation with backing from AWS, Google Cloud, Oracle, and Ericsson, carrying forward the last BSD-licensed codebase. Redis then shifted to AGPLv3 with Redis 8 in 2025, but the trust cost was already paid — surveys after the change found a large majority of Redis users actively evaluating alternatives, and the hyperscalers have kept investing in Valkey rather than walking back. Practically: the two are API-compatible for most workloads, and the functional gap is small (Redis 8 adds things like native vector sets). The real decision is governance, not features. If BSD licensing and open governance matter to your legal team, Valkey. If they don’t, either works. Whichever you pick, its job is the same — sit in front of your primary store and absorb the reads that would otherwise flatten it.

Workload 3: Answering questions (cloud and analytics)

Snowflake

Separating storage from compute was the structural idea that made Snowflake work: you pay for what you query, not what you hoard. It reads structured tables, JSON, and Parquet in the same query without a pipeline in between. The underrated feature is the Data Marketplace — pulling third-party feeds and financial datasets in without an ETL job or a separate vendor contract. For finance teams, that’s the difference between monthly reporting and something closer to live. Choose it when: analysis across millions of rows is a daily activity, not a quarterly one.

Firebase and Supabase

For a two-person team that needs to be in production next month, a managed backend is the whole answer. Both bundle database, auth, real-time sync, and storage into one platform. Supabase’s advantage is structural: it’s Postgres underneath. Your queries stay portable, so outgrowing the free tier means moving data — not rewriting the application. Firebase is faster to start with and harder to leave.

Three rules that separate mature architectures from the rest

1. Match the engine to the job. Transactions want ACID. Analytics wants columnar storage. Real-time wants memory. Using one system for all three doesn’t fail loudly — it degrades quietly, which is worse.
2. Plan for horizontal scale before you need it. Bigger machines have a ceiling; more machines don’t. Postgres, MongoDB, and CockroachDB shard and replicate natively. Retrofitting that into a live monolith is one of the genuinely miserable engineering projects.
3. Separate PII from operational data by design. Sensitive customer data belongs in an access-restricted store, joined to transactions through anonymized IDs. Done right, a breach of your transaction database reveals that User_8492 moved $10,000 — and nothing about who that is. GDPR and PCI-DSS treat this as a requirement, not a suggestion.

Do you need something custom?

Almost certainly not. Before anyone writes a proposal for a bespoke engine, exhaust the extension ecosystem — Postgres alone covers time-series (TimescaleDB), AI embeddings (pgvector), and geospatial (PostGIS) without a line of engine code. Custom is expensive to build and roughly twice as expensive to keep alive.

What nobody puts in the pitch deck

Knowing which database you should be on is the easy half. Getting your data off a fifteen-year-old Sybase box, an aging on-prem SQL Server, or an inherited MySQL estate is where modernization actually stalls — usually on the embedded business logic nobody documented and everyone depends on. That work is more tractable than it used to be. Tooling now automates schema conversion, stored-procedure translation, and query rewriting across the common source-target pairs; migration specialists like Ispirer report automation rates in the high nineties on Oracle-to-PostgreSQL projects, which turns a multi-year rewrite into something a team can plan around. It’s worth scoping that before you assume a migration is off the table on cost grounds.

In short

• Transactions → PostgreSQL, or Oracle if regulation demands it
• Microsoft shop → SQL Server
• Read-heavy and simple → MySQL
• Flexible documents → MongoDB
• Sub-millisecond reads → Valkey (open governance) or Redis (features)
• Analytics at scale → Snowflake
• Small team, fast launch → Supabase (portable) or Firebase (fastest)

Every one of these choices compounds. The database that fits your workload today is mostly valuable because of the migration it saves you from in 2029.