Databases compared: pricing, serverless, vector and full-text search
Buyers land here holding a credit rate next to a flat monthly plan, with no way to line the two up. We priced 14 engines and platforms across 54 plans and put every unit into dollars.
How much does a database cost, and what unit is that price in?
Where a monthly figure exists at all, it starts near $4.99 and stops near $1,800, and the halfway mark is $82.76. Prices checked July 29, 2026. Everything else prices in a unit of its own: a credit, a DBU, a Request Unit, an hour of compute. Those convert. The multiplier lives in the plan, not the unit.
The vendors that meter usage here run on eight distinct units: credit, DBU, Request Unit, vCPU-hour, ECPU-hour, VCU-hour, instance-hour, rows read. All eight are published. Disclosure is not the hard part on this market.
One unit can carry more than one price. A Snowflake credit costs more on each edition step up, and a Databricks DBU is priced by the kind of work it runs, from AI jobs at the bottom of the card to interactive sessions at the top.
Compute is one meter. Storage is another, metered separately almost everywhere. Egress appears on the platforms that charge for data leaving. Databricks bills only its own share, and the cloud provider underneath invoices the virtual machines.
Most of this page starts free. That is 12 vendors from 14, and on the open-source engines the zero is not a tier at all. It is the whole product, and what you pay for sits outside the vendor's price list entirely.
An empty price cell usually points at a rate card. It rarely points at a secret. Sealed tiers are the smaller story here: Nine of them across 54 plans we read, and most of them are a top step nobody can buy without asking. The rest sit lower down the range.
Ranked by a transparency score: pricing transparency 60%, user satisfaction 40%. Capability is not scored. It is the condition grid below. Prices are read from vendor pricing pages and re-checked per product on the dates shown. ComparEdge sells no databases and takes no payment for placement. How the score is built.
01 / 06
Databases ranked: pricing transparency, serverless, vector and ACID
Rows run in score order. Only a vendor with a fixed price per month reaches the price axis. A credit rate has no monthly point to occupy, and those rows carry a rate-card marker instead.
Sorted by transparency scorePriced tiers 45 / 54Full disclosure 6 / 14How to read this table
Reading the console
The axis plots one figure per vendor, the cheapest plan with a fixed price per month. A vendor selling credits or compute-hours has no such plan, and its row shows a rate-card marker in place of a tick. Free rows are the engines with nothing to sell you. Neither kind is withholding a number. The condition columns measure breadth, not quality. A cache and a distributed SQL store solve unrelated problems, and where a cell is empty we found no such capability at verification time. Read Serverless with care. On some rows it names a billing model. On others it names a managed service a cloud provider sells around the engine.
Price axis
One shared logarithmic axis, $4.99 to $2k. A tick further right is genuinely dearer. The tall tick is the cheapest paid tier.
Diamond
A flat platform fee, billed per month, not per seat.
Granted
Listed on the vendor’s own pages on the verify date.
Conditional
Present, but some tiers carry no published price.
Not on the record
Absent from the recorded feature set. It means unlisted, not incapable.
01DuckDBFree, no paid plansFree tierAll tiers pricedOpen source listedServerless not on the recordACID txns listedVector search listedFull-text search not on the recordTime-series listedFree100Alternatives to DuckDB
02PostgreSQLFree, no paid plansFree tierAll tiers pricedOpen source listedServerless listedACID txns listedVector search listedFull-text search listedTime-series not on the recordFree91Alternatives to PostgreSQL
03DatabricksUsage-priced, public rate cardNo free tierAll tiers pricedOpen source listedServerless listedACID txns listedVector search listedFull-text search not on the recordTime-series not on the recordSales only91Alternatives to Databricks
04MySQLFree tierAll tiers pricedOpen source listedServerless listedACID txns listedVector search not on the recordFull-text search listedTime-series not on the record$178.33/server88Alternatives to MySQL
05Oracle DatabaseUsage-priced, public rate cardFree tierAll tiers pricedOpen source not on the recordServerless listedACID txns listedVector search listedFull-text search listedTime-series not on the recordSales only88Alternatives to Oracle Database
06SurrealDBUsage-priced, public rate cardFree tier1 of 3 tiers unpricedOpen source listedServerless not on the recordACID txns listedVector search not on the recordFull-text search listedTime-series listedSales only85Alternatives to SurrealDB
07OpenSearchFree, no paid plansFree tierAll tiers pricedOpen source listedServerless not on the recordACID txns not on the recordVector search listedFull-text search listedTime-series listedFree84Alternatives to OpenSearch
08TiDBFree tier1 of 4 tiers unpricedOpen source listedServerless listedACID txns listedVector search not on the recordFull-text search not on the recordTime-series not on the record$1,800flat83Alternatives to TiDB
09SnowflakeUsage-priced, public rate cardNo free tier1 of 4 tiers unpricedOpen source not on the recordServerless not on the recordACID txns listedVector search listedFull-text search not on the recordTime-series not on the recordSales only82Alternatives to Snowflake
10ClickHouseFree tier1 of 4 tiers unpricedOpen source listedServerless not on the recordACID txns not on the recordVector search listedFull-text search not on the recordTime-series listed$66.52flat81Alternatives to ClickHouse
11TursoFree tier1 of 5 tiers unpricedOpen source listedServerless listedACID txns listedVector search not on the recordFull-text search not on the recordTime-series not on the record$4.99flat80Alternatives to Turso
12RedisFree tier1 of 5 tiers unpricedOpen source listedServerless not on the recordACID txns not on the recordVector search listedFull-text search listedTime-series listed$5flat79Alternatives to Redis
13CockroachDBUsage-priced, public rate cardFree tier1 of 4 tiers unpricedOpen source listedServerless listedACID txns listedVector search not on the recordFull-text search not on the recordTime-series not on the recordSales only79Alternatives to CockroachDB
14ElasticsearchFree tier2 of 7 tiers unpricedOpen source not on the recordServerless listedACID txns not on the recordVector search listedFull-text search listedTime-series listed$99flat74Alternatives to Elasticsearch
GrantedSome tiers sealedNot on the recordCheapest paid seatFlat account fee, not per seatScore ranks pricing transparency and user ratings, not capability. Capability is the grid.
Scroll the console sideways to reach the remaining conditions.
02 / 06
Every database's plans, verdicts and the date we checked each price
Position on this list is our score, nothing more. The list mixes engines you install yourself with platforms that rent you time on theirs, and buyers really do cross that line when they shortlist. Expand any row to see every tier and the unit each is billed in.
Analysts who want to query Parquet and CSV files inside a Python or R session with no server to stand up. DuckDB is free with no commercial tier above it, and it stays single-node by design: the day the data outgrows one machine, so does the tool.
Critical gapSyntax differences create friction for engineers migrating from BigQuery.
Plan table and expert take
DuckDB: expert take
The limit arrives early: one machine, one process, no concurrent writers. DuckDB sells nothing at all, and there is no commercial tier above the free one to grow into. Outgrowing it means a migration, and that is worth planning on day one.
Where DuckDB holds up
Runs in-process with zero infrastructure setup
Directly queries Parquet and CSV on S3 without ETL
Outperforms many server-based DBs on single-machine workloads
Teams wanting a single relational engine they can extend into JSON, full-text and vector work instead of running three systems. The published price is zero and stays zero. The bill arrives as tuning and as the operations Postgres will not do for you.
Plan table and expert take
PostgreSQL: expert take
Teams who read free as cheap meet the bill later. Postgres publishes zero and means it, and the money moves to hardware, upgrades and whoever handles the tuning the engine will not do for itself. Managed Postgres is a different vendor's purchase.
Where PostgreSQL holds up
Extensible architecture supports GIS, time-series, and vector data
ACID compliance and MVCC ensure high data integrity and concurrency
Data teams that have reached the stage where analytics and machine learning must share one store rather than two. Databricks sells the lakehouse itself, not a database you slot into an existing stack, and it opens with nothing free to test on.
Plan table and expert take
Databricks: expert take
Buyers asking for the DBU price get a rate card, not a number: every kind of work carries its own rate, from AI jobs up to interactive sessions. Photon nearly triples what a job burns. The virtual machines underneath land on the cloud provider's bill.
Where Databricks holds up
Lakehouse architecture eliminates ETL between data lake and warehouse
MLflow is the de-facto ML experiment tracking standard
Unity Catalog provides unified governance across data and AI
Web teams long settled on MySQL who want the familiar engine with vendor support behind it. Community stays free. The commercial editions are Oracle subscriptions counted per server on an annual basis, which turns them into a procurement decision rather than a signup.
Critical gapThe native version control workflow necessitates manual synchronization between files and Git.
Plan table and expert take
MySQL: expert take
Shops running Community rarely meet the other MySQL. Oracle sells the supported editions on annual subscriptions counted per server, with Standard Edition at $178.33 a month on that basis and two heavier editions above it.
Where MySQL holds up
Most widely deployed database: abundant expertise and tooling
Enterprises consolidating relational and vector work onto an engine they already license. Oracle meters the cloud editions by the ECPU-hour and runs an unusually generous free tier underneath. An existing licence cuts the hourly rate to a fraction of the list one, which is where the economics really get decided.
Critical gapThe absence of 23ai on-premises support delays migration from 19c.
Plan table and expert take
Oracle Database: expert take
The conversion trap comes first: an OCPU is two vCPUs, and an ECPU-hour never lines up with another vendor's hour untouched. Bringing your own licence cuts the rate to roughly a quarter. Storage, backup and egress each run a meter of their own.
Where Oracle Database holds up
45+ year track record in mission-critical enterprise deployments
Developers early enough in a project to bet on document, graph and relational models in one engine with live queries. SurrealDB gives one instance and a gigabyte free with no expiry, then meters by the hour that instance runs. The managed cloud is young, and that is the risk you are taking on.
Critical gapProduction users report memory spikes and data-corruption issues.
Plan table and expert take
SurrealDB: expert take
The young-product risk is priced into everything else here. SurrealDB gives one instance and a gigabyte free with no expiry, then meters instance-hours, which makes a prototype genuinely free and a production cluster a moving number. The enterprise step is quoted, not published.
Where SurrealDB holds up
Multi-model flexibility: document, graph, and relational in one
Built-in auth eliminates separate auth layer for simple apps
Teams deciding whether to run search themselves or rent it from somebody else. OpenSearch costs nothing at any size, with the licence and the feature set open. What you pay for is the cluster: sizing it, keeping it healthy, and the night somebody has to fix it.
Critical gapUnbalanced ingestion creates shard hot-spotting and JVM memory overhead.
Plan table and expert take
OpenSearch: expert take
Teams pricing OpenSearch usually end up pricing somebody else. The engine costs nothing and has no paid tier at all. The managed service people actually buy comes from a cloud provider on that provider's rate card, and the bill is the provider's.
Where OpenSearch holds up
Apache 2.0 license - no SSPL restrictions
API-compatible with Elasticsearch for easy migration
Built-in security plugin at no extra cost vs Elastic
Teams whose single relational instance has hit its ceiling and now need horizontal scale with transactional and analytical queries in the same place. The free tier on TiDB is unusually large before anything is owed. The distributed design is overkill below that line, where the hardware it wants outweighs what it returns.
Plan table and expert take
TiDB: expert take
Teams sizing TiDB usually start free: 250 million Request Units and two 25 GiB stores a month, with per-unit overage after that. Above the free tier the unit changes to provisioned capacity, which is a different purchase with a different shape of bill.
Where TiDB holds up
HTAP eliminates separate analytics warehouse for many use cases
MySQL compatibility reduces migration complexity
Strong horizontal scaling for high-write workloads
Free serverless tier for development
★★★★★4.7CE scoreG2 4.6 · 67 reviewsFounded 2015Verified July 8, 2026
Analytics teams that want storage and compute to scale independently and have somebody watching the spend. Snowflake runs no free tier. Warehouses bill only while they run, which makes a suspended cluster free and an unwatched one the fastest route to a bill nobody predicted.
Critical gapCompute costs scale aggressively with inefficient query execution.
Plan table and expert take
Snowflake: expert take
The trap is that a credit is not a price. Snowflake charges $2 per credit on Standard and $4 on Business Critical, and the same query costs twice as much one edition up. Rates are quoted for one reference region.
Engineers running sub-second analytical queries over very large columnar tables. ClickHouse is free to self-host, and its managed tier prices monthly from Basic at $66.52. Joins and mid-stream schema changes are where the engine pushes back.
Critical gapOpen source deployments require extensive manual pipeline engineering for reliable data ingestion.
Plan table and expert take
ClickHouse: expert take
The extra meters are where this one surprises people. Compute and storage bill on separate lines, data leaving is charged by the gigabyte, and pulling data in through ClickPipes carries its own rate. None of that shows in the plan price.
Where ClickHouse holds up
Scans billions of rows per second via vectorized query execution
Materialized Views for real-time, incremental data aggregation
High-ratio data compression (LZ4, ZSTD) drastically cuts storage costs
Natively handles semi-structured data (JSON, maps, arrays) at scale
Full-featured, production-ready open-source core; no vendor lock-in
Teams building multi-tenant or edge apps that have reached the stage where one database per customer starts to make sense. Turso counts rows read rather than hours, and the free plan carries half a billion monthly reads. It trades away the heavier relational features to get there.
Critical gapBase tier transaction latency remains at 100 ms.
Plan table and expert take
Turso: expert take
The overage model is where the bill goes sideways. Turso meters four things at once: storage, rows read, rows written and syncs, each with its own rate. Overages must be switched on. Leave them off and you hit a wall instead of a charge.
Where Turso holds up
SQLite familiarity with global edge replication
Database-per-tenant trivially cheap for multi-tenant SaaS
Applications needing sub-millisecond caching or pub/sub next to the main store. Redis is free to self-host, and its managed service is quoted hourly instead of per plan. It earns a place as a companion, not as the system of record.
Critical gapThe RAM-based architecture restricts storage for massive datasets.
Plan table and expert take
Redis: expert take
Teams moving up to Pro should check the cloud list first: Pro sells on two clouds where the cheaper tiers offer three. Redis gives the first $200 on Pro free, which is exactly one month at that tier's minimum spend.
Where Redis holds up
Sub-millisecond latency: fastest data store for caching
Universal: supported by every framework and language
Rich data structures for real-time use cases
Redis Stack adds search, JSON, and vector capabilities
Products that have to keep transactional SQL correct across regions without hand-managing replication. CockroachDB starts on Request Units and moves to provisioned vCPU-hours on the tiers above. The unit changes as you grow. Reporting-heavy work will feel the mismatch: this is built for resilience, not analytics.
Plan table and expert take
CockroachDB: expert take
Teams starting on Basic get 50 million Request Units and 10 GiB of storage free each month, with signup credits on top and no card required. The step up swaps the unit for provisioned vCPU-hours. Five nines applies only to a multi-region Advanced deployment.
Where CockroachDB holds up
Survives disk, machine, rack, and datacenter failures with zero data loss.
Horizontally scales both reads and writes by simply adding new nodes.
PostgreSQL wire-protocol compatibility simplifies migration and tooling.
Geo-partitions data to keep it close to users, reducing latency.
Offers strongly consistent, ACID-compliant transactions across regions.
Search and log analytics teams after a mature engine, willing to size a cluster to keep it fast. Elasticsearch is free to self-manage, and the paid subscriptions on that route are sold by quote rather than published.
Critical gapSelf-managed Kubernetes deployments require high administrative overhead for cluster maintenance.
Plan table and expert take
Elasticsearch: expert take
The published tier prices are starting points, not prices. Elastic quotes hosted tiers from $99 on Standard for a small deployment, and a real cluster is metered by memory over time. The serverless track uses another unit again, the VCU-hour.
Where Elasticsearch holds up
Best-in-class full-text search with relevance tuning
Rich aggregation engine for log analytics and dashboards
Massive ecosystem with Logstash, Kibana, and Beats
★★★★★4.6CE scoreG2 4.5 · 292 reviewsCapterra 4.6Verified July 8, 2026
Database questions: monthly cost, per instance or per database, credits
How much does a managed database cost per month?
Wide, and the width is real. Among vendors with a flat plan the middle sits at $82.76 as of July 29, 2026, and the two ends are far enough apart that the middle says little by itself. Everything else bills by consumption, where the number follows the hours you keep compute running rather than the tier you picked. Size the workload first.
Is the bill for the instance, or for each database on it?
Both models sit on this page. Provisioned platforms charge for the instance: CockroachDB meters vCPU-hours on its Standard plan whether one database or ten live on it. Consumption platforms charge for the work instead. TiDB counts Request Units and Turso counts rows read, and on those a second database costs nothing until it does something.
What am I paying for in a Snowflake credit or a Databricks DBU?
An hour of somebody else's compute, wrapped in a name. A Snowflake credit buys warehouse time, and that same credit costs more on the higher editions. A Databricks DBU buys processing priced by the kind of job, with AI work at the bottom of the rate card and interactive sessions at the top. Neither price includes storage.
Why are managed databases so expensive next to a free engine?
You are buying operations, not software. PostgreSQL and OpenSearch hand you the engine and nothing else, and every hour of running it lands on your side. A managed platform takes that work and prices it into the unit rate, with margin on top. The gap narrows once a platform team already exists. It rarely closes for a small one.
Does a database that sits idle still cost money?
On several here, no. Snowflake charges credits only while a warehouse runs, and a suspended one costs nothing. ClickHouse scales an idle service to zero and does not bill the pause. CockroachDB Basic scales to zero the same way. Provisioned tiers work the other way round. You reserved the hardware, and it bills whether queries arrive or not.
Is there a floor under an hourly database rate?
Almost always, and it decides a small bill. Redis quotes Essentials by the hour, then applies a $5 monthly minimum underneath it, and the Pro tier floors at $200. Provisioned platforms have a softer floor: the smallest cluster they will let you run. An hourly rate with nothing under it is the exception on this page.
Which meters run alongside the compute charge?
Storage, first and nearly everywhere. Snowflake prices storage by the terabyte-month on top of credits, and ClickHouse bills it on a separate line from compute. Egress is the next one, charged on data leaving. Ingestion can be a third. Databricks adds a fourth: the cloud provider underneath invoices you for the virtual machines it ran.
Where does serverless stop and a provisioned cluster start?
At the moment you begin paying for capacity you are not using. Serverless tiers meter work: Elasticsearch bills VCU-hours, TiDB counts Request Units, and both fall toward zero when traffic does. A provisioned cluster reserves vCPUs by the hour and bills them whether queries arrive or not. Watch the word, though. On some engines it names a managed service a cloud provider sells, not a billing model.
What does it cost to run a free engine yourself?
The licence is the only zero in it. You provision capacity ahead of demand instead of consuming it, and the hardware bill does not fall when traffic does. Storage and replicas are yours to size. Upgrades, backups and tuning are standing work for whoever owns the cluster. MySQL shows the other half: Community costs nothing, and Oracle sells the supported editions as a per-server subscription on top.
What do the top tiers add that the paid entry tier does not?
Isolation and paperwork, mostly. An entry paid tier tends to be shared infrastructure with a best-effort SLA. Above that you get dedicated networking, keys you control and an uptime commitment worth quoting. ClickHouse Enterprise adds private regions and HIPAA, and Redis Pro adds a dedicated VPC with active-active replication. Capacity limits move as well: an entry tier can cap storage where the step above does not.
Which unit should decide a shortlist?
Workload shape decides it, not the rate itself. Spiky and occasional favours anything metered, whether that is credits, Request Units or rows read, which fall to nothing between jobs. Steady and constant favours a provisioned hour or a flat plan, where the rate is lower and the meter never stops. Get that backwards and the same workload runs several times dearer with no price changing anywhere.
Is every product here really a database?
Not in the same sense, and the price axis will mislead you if you forget it. A cache, a search engine, a warehouse and a transactional store all sit on this page. Teams genuinely shortlist across them. Redis holds hot keys in memory and ClickHouse answers analytical queries over huge tables, and a monthly figure from one says nothing useful about the other.
Field note 01
Making two database prices comparable
Start with a workload, not a price. Pick the hours you expect compute to be busy, the data you intend to keep, and the reads you will run against it. Those numbers are the only common ground this market has. Every vendor's unit converts through them.
Then apply the multiplier that vendor uses, which is the part buyers skip. A credit rate moves with the edition. A DBU rate moves with the kind of job, and switching Photon on raises what the same job consumes by close to three times. Add storage, metered separately almost everywhere. Add the floor if the tier has one. What comes out is a monthly number, and it is the first thing here that means the same on two different price lists.
Field note 02
An empty price column is not a hidden price
A blank on the axis has several causes here, and they point in opposite directions. An open-source engine shows nothing for the simple reason that it is free. A consumption platform shows nothing for the opposite one: it publishes a rate card in place of a plan, and a rate has no monthly point to sit on. The ask-us case is narrower than it looks. Nine tiers across 54 plans carry no number, and most of those are a last rung rather than a way in.
The reverse trap is a marker that looks like a door. The highest monthly figure here, $1,800, is TiDB's provisioned Premium cluster. Its free Starter tier and its hourly Dedicated rate have nowhere to plot, and the axis shows the one number with the right shape. Read the top of that range as a cluster somebody sized, not as the cost of getting in.
The verdict on databasesSigned review · Updated
Oleh KemFounder & Lead AnalystComparEdge Editorial
Nothing here compares until you know the unit. Every figure on this page is quoted in a currency its vendor invented, and two of them using the same word rarely mean the same money.
Two questions settle the rest. Find the meter running beside compute, because storage almost always bills apart from it. Then ask whether an hourly rate has a monthly minimum sitting under it, which several of these do. A licence fee of zero earns a high number here. That is a statement about disclosure and never a claim that the engine is cheap to keep running.
MethodWe read 54 plans across fourteen engines, converted every vendor currency we found and dated each figure .
DisclosureCollection is tool-assisted; every verdict is written and signed by a human analyst.
05 / 06
Not sure which? Answer one, take a shortlist.Pick the line that sounds like your team. Each one opens the vendor built for it.
How this review is made. Prices are read from vendor pricing pages and re-checked on the dates shown against each product. Condition columns reflect the feature set recorded on the vendor’s own pages on that date. ComparEdge sells no databases and takes no vendor payment for placement. Where a vendor publishes nothing, this page says so rather than estimating. Ranking is by transparency score: pricing transparency 60%, user satisfaction 40%. What a product can do is shown in the condition columns and carries no weight in the number.
Oleh Kem
