Filament vs Fivetran: Open Source vs Managed ELT

Filament and Fivetran differ on who runs the pipeline and who pays per row, and that decides the choice. Fivetran is proprietary managed ELT billed on monthly active rows from a published rate card, with the pipeline running in Fivetran's cloud or through a hybrid agent. Filament is an Apache 2.0 replication engine you run yourself with no meter. As of October 2026, pick by operations capacity and connector coverage.

What you pay for Filament vs Fivetran

Fivetran's bill scales with distinct rows changed and Filament's with compute hours, so churn decides the crossover. Fivetran defines the unit plainly, "MAR are unique identifiers, or primary keys, that we use to track transfers from your source system to your destination each month" (Fivetran usage docs). A row counts once a month however often it changes, and initial syncs are free. Deletes count on pay as you go accounts since January 1, 2026 (Fivetran 2026 pricing updates).

The rate card is public, which the top ranking pricing explainers often miss. The Service Consumption Table lists a base spend plus an incremental spend per million MAR at each threshold, applied per connection with no pooling across connections (Fivetran Service Consumption Table).

The formula is Base Spend per Month + (actual MAR - threshold) * Incremental Spend per Million MAR. Annual contracts discount it by 5 to 22.6 percent on Standard with a 12,000 dollar minimum (Fivetran billing and plans).

Filament has no rate card because there is no charge. The announcement post states the position, "Same engine, no gates, nothing held back for a paid tier" (Introducing Filament). Galaxy builds Filament.

What replaces the meter is a footprint you size yourself. A Filament deployment needs a Postgres datastore, a NATS event bus, and a secret provider, and in Kubernetes mode "creates one worker Job per run" (Filament deployment overview). Warehouse credits are paid under both tools, since Fivetran itself recommends an X-Small Snowflake warehouse with one minute auto suspend (Fivetran Snowflake destination).

The dollar column is list price computed from Fivetran's published formula before any discount, not a quote. The last two rows are the honest case for Fivetran. "Unchanged rows in re-import tables" are free MAR (Fivetran usage docs), while a Filament full reload pays compute for every scan.

The keyless row is the trap, because in a table without a primary key "Each row in the table counts towards MAR".

Where does the data get processed?

Fivetran processes data in its own cloud or through an agent you host, and Filament processes it only where you run it. Fivetran's SaaS model works "entirely within the Fivetran cloud", and its Hybrid Deployment runs an agent in your network while "Only metadata (including MAR information) and logs are sent to Fivetran" (Fivetran Hybrid Deployment).

Hybrid needs Enterprise or Business Critical and an x86 host sized at 8 vCPUs and 32 GB of memory for one or two connections, with no ARM support (Fivetran hybrid setup guide).

Filament has no vendor plane at all. It runs as "a standalone binary, deploy it on Kubernetes, or embed it in a Go application" (Filament introduction), and the deployment guide warns to "Run exactly one control-plane replica" (Filament deployment overview). Nothing leaves your account unless you point OpenTelemetry at a collector.

The sync overview says "You must have an Enterprise or Business Critical plan to use the 1-minute sync frequency" (Fivetran sync overview). Private networking is a Business Critical feature (Fivetran features), and the response times come from the support policy effective June 23, 2026 (Fivetran support policy).

How each captures Postgres changes

Both read the write ahead log through pgoutput, and they part ways on the fallback and on deletes. Fivetran's Postgres connector offers logical replication or a Query-Based method on the xmin column (Fivetran Postgres connector). Under the query path, "If capturing deletes is disabled" Fivetran "will not recognize deleted rows at all".

Deletes under logical replication arrive as a soft delete flag, and the same page warns that xmin wraparound can trigger re-syncs and MAR spikes.

Filament creates the slot with an exported snapshot so "the baseline and the change stream share one consistent point" (Filament Postgres source). It splits large tables into as many as 64 parallel ranges and requires wal_level=logical plus a primary key on every CDC table. CDC merge applies deletes as deletes, CDC append keeps them as events, and there is no xmin fallback.

Schema handling is add only on Filament. Fivetran applies column type changes under logical replication and still needs a re-sync for drops and renames.

Integrity and recovery

Filament checks bytes on every batch, and Fivetran contracts outcomes. Before each write Filament "calculates a CRC32-C checksum over the batch's Arrow buffers, its row and column counts, and its insert/update/delete operations" and fails the run on a mismatch (Filament integrity and checkpoints). Recovery is "deliberately at-least-once" with durable checkpoints. The same page concedes these checks "do not universally read data back" from the destination.

Fivetran documents no per batch checksum. Its answer is operational, with resumable initial syncs that copy "a limited number of rows at a time" and table level re-syncs. The contract is an SLA of 99.9 percent on schedule delivery on Enterprise and Business Critical, with credits capped at half the fee for the affected connection (Fivetran SLA).

Connectors and maturity

This is where Fivetran wins outright and Filament must be read with its labels on. Fivetran markets "700+ fully managed connectors" (Fivetran pricing), many of them Lite connectors that "often cover fewer endpoints" but carry the same SLA (Fivetran Lite connectors).

Filament lists 28 sources and 14 sinks. Its maturity page defines alpha as "Documentation-derived, not yet tested end to end" and beta as "Runs end to end. Not verified for every supported object or environment" (Filament connector overview).

Eight sources and the Postgres sink are beta, the Snowflake, BigQuery, Redshift, and ClickHouse sinks are alpha, and no production connector is labeled stable. New SaaS sources are YAML manifests over one HTTP connector, so adding an API is a manifest rather than code.

Where Fivetran wins

Fivetran wins on breadth, paperwork, and the long tail of operations. It holds SOC 1 and 2, ISO 27001, a HIPAA BAA, and PCI DSS Level 1 and HITRUST on Business Critical (Fivetran security), and it processes data in seven regions. Hybrid Deployment and HVR, "a self-hosted version of Fivetran" for heavy database replication (Fivetran HVR docs), cover regulated estates that still want a vendor on call.

It also ships features Filament lacks today. History mode gives type 2 slowly changing dimensions on every Postgres table, keyless tables sync through a synthetic key, and a query based fallback covers databases where logical replication is impossible. Transformations run dbt models with 5,000 model runs free a month, and the dbt Labs merger closed on June 1, 2026 (Fivetran press release).

A mostly static large table also costs Fivetran customers almost nothing, while Filament scans it on every full run.

Which tool fits which job?

Run Filament when you have a database source, a warehouse sink, and an engineer who can operate a Helm chart. Buy Fivetran when the source list is long or nobody will run software.

Related guides

• Best Airbyte Alternatives in 2026 ranks the wider field including Fivetran.

• Which Database Replication Tools Are Actually Open Source checks the LICENSE files behind the labels.

• Postgres to Snowflake, the complete replication guide is the route most readers of this page are pricing.

• Full Load vs Incremental vs CDC explains the read modes behind the MAR arithmetic.

• Database replication, the complete guide is the pillar this comparison hangs off.

• Introducing Filament is the announcement with the deployment options.

Frequently asked questions

What is a monthly active row in Fivetran pricing?

A monthly active row, or MAR, is a distinct primary key that Fivetran inserted, updated, or deleted in a calendar month, counted separately per connection and table. A row that changes ten times in a month counts once. Initial syncs are free. Since January 1, 2026 deletes count toward paid MAR on pay as you go accounts.

Does Fivetran publish its price per million MAR?

Yes. Fivetran's Service Consumption Table, effective February 1, 2026, lists a base spend and an incremental spend per million MAR at each volume threshold for the Standard, Enterprise, and Business Critical plans. The first million MAR on Standard lists at 495 dollars, and the rate per million falls as a single connection grows.

How does a full re-sync count toward Fivetran MAR?

Only changed rows count. Fivetran's usage docs walk through a re-import table whose unchanged rows cost zero MAR. In the next sync only the new and changed rows are billed, 40 MAR out of a 120 row re-import in their example. Initial historical syncs and most re-syncs are free.

Is Filament free to use?

Filament is Apache 2.0 with no paid tier, no per row charge, and no hosted plan published, so the license cost is zero. You still pay for the machines that run its server, control plane, Postgres datastore, and workers, plus the warehouse credits any loader consumes. Zero license cost is not zero cost.

Which Fivetran plan do you need for 1 minute syncs, Hybrid Deployment, or PrivateLink?

One minute syncs and Hybrid Deployment both require the Enterprise or Business Critical plan, per Fivetran's sync overview and deployment docs. Private networking through AWS PrivateLink, Azure Private Link, or GCP Private Service Connect requires Business Critical, which also carries customer managed keys and region selection.

Where is data processed in Fivetran compared with a self hosted tool?

In Fivetran's SaaS model data is processed entirely inside Fivetran's cloud, in one of seven regions from the US and EU to India and Singapore. In Hybrid Deployment an agent in your network processes data while Fivetran still runs the control plane and receives metadata and MAR counts. A self hosted engine keeps both in your account.

Is Filament a Fivetran alternative for Postgres to Snowflake change data capture?

Yes for a team that can run Kubernetes, with two caveats. Filament reads Postgres through a pgoutput replication slot and writes typed Snowflake tables from staged Parquet with no per row meter. Its Postgres source is labeled beta and its Snowflake sink alpha, and no public benchmark covers Fivetran, so this comparison makes no speed claim.

Does Fivetran have an uptime or delivery SLA?

Yes, on paid plans. Fivetran guarantees 99.9 percent availability of its core services and, on Enterprise and Business Critical, 99.9 percent on schedule delivery once a table has synced for five consecutive days. Service credits are capped at half of the monthly fee for the affected connection and are the sole remedy.

What happened between Fivetran and dbt Labs?

Fivetran announced an all stock merger with dbt Labs on October 13, 2025 and completed it on June 1, 2026. George Fraser stayed on as CEO and Tristan Handy became President. Fivetran's Transformations product runs dbt models, with a free monthly allowance of model runs and published per run rates above it.

How does Filament handle schema changes compared with Fivetran?

Filament applies add only schema evolution. It creates new columns with ADD COLUMN IF NOT EXISTS and never drops, renames, or retypes a column, with strict and nullable add policies on the Postgres sink. Fivetran's automatic schema migration adds columns and, under logical replication, applies column type changes, while drops and renames still require a re-sync.