Zipline vs. Databricks, Snowflake, Chalk, and Feast
How Zipline compares with Chalk, Databricks, Snowflake and Feast on iteration speed, scale, serving cost and training-serving consistency.
Every real-time feature platform now lets you define a feature once and use it for training and serving. This post compares Zipline with Chalk, Databricks, Snowflake and Feast on what matters once a model is in production.
In short: on Zipline, a new 30-day feature serves correct values within hours. The others wait days for streaming to fill the window, or reprocess the feature's whole history first. Each lookup is one key read, so at 100,000 requests per second Zipline costs 5–9x less to serve than Chalk and Databricks.
At a glance
| Zipline | Chalk | Databricks | Snowflake | Feast | |
|---|---|---|---|---|---|
| New 30-day feature ready | Hours | After backfill | ~2 days | Full history recompute | 30 days |
| Each lookup | 1 key read | Merges buckets | Merges rows | Postgres query | 1 row read |
| Serving 100k req/s | 1 Valkey cluster | 1,250 vCPUs | 67 stores | 800 nodes | DIY |
| Monthly cost, 100k req/s | $48.4k | $260k | $443k | $753k | DIY |
| Consistency check | Built in | No | No | No | No |
| Generally available | Yes | Yes | Oct 2026 | Oct 2026 | Pre-1.0 |
| Self-host | Via Chronon | No | No | No | Yes |
Costs assume 20 million entities and 20 feature groups per request, 5 of them streaming, with high availability everywhere. The calculator at the end lets you change all of that.
New features serve correct values in hours
Zipline seeds a new window from history in its first batch upload, so a 30-day feature is correct online as soon as that one job finishes. The others wait for the stream to fill the window, for a full history recompute, or for someone to run a backfill.
Show as table
| Platform | Each new training set | Changing a feature |
|---|---|---|
| Zipline GigaTile | Point-in-time correct backfill from the same definition, reusing cached daily partial aggregates. | Reuses feature groups whose definitions have not changed and recomputes the rest. |
| Chronon open source | Point-in-time correct backfill from the same definition. | Rerun the changed feature group’s backfill and upload. |
| Databricks RollingWindow | Streaming and rolling windows are never stored offline, so every training set recomputes them from raw events. | No update method: create a new feature and materialize it again. |
| Databricks SawtoothWindow | Recomputed from the stream’s ingestion table for every training set. | No update method: create a new feature and materialize it again. |
| Snowflake feature views | ASOF joins over the feature view’s dynamic table; tiled windows need an option marked “internal use only”. | Definitions are immutable: a new version recomputes the full history. |
| Chalk materialized buckets | Aggregates raw events for each row by default; bucket mode builds buckets once per query. | A changed feature starts empty under a new key until it is backfilled. |
| Feast your pipelines | Latest stored row at or before each label’s timestamp (a snapshot lookup). | Change and rerun your own pipelines. |
Edits work the same way. Zipline recomputes only the feature groups whose definitions changed. Snowflake definitions can't be changed, so every edit recomputes the full history, and Databricks has no update method for a registered feature.
One key read per lookup
What a platform stores for a window decides how much it reads on every request, and that limits how far it scales. GigaTile, which we built with OpenAI and Etsy, stores one finished value per entity. The other layouts store raw events or many partial values and merge them at read time.
Show as table
| Layout | Entries per entity | Work per lookup | Oldest edge of the window |
|---|---|---|---|
| Raw events Databricks RollingWindow · Chalk default | 3,000 | range scan of 3,000 events, aggregated at read time | exact |
| Hops Chalk materialized buckets · Chronon tiled mode | 720 | range scan of 720 hourly buckets, merged at read time | bucket size (oldest bucket counts in full) |
| Sawtooth Chronon open source · Databricks SawtoothWindow | 49 | 1 read + scan of 48 recent hops, merged at read time | Chronon: 1 day hops · Databricks: daily, 18:00 UTC |
| GigaTile Zipline | 1 | 1 key read, combined at write time | 1 day hops |
- Zipline reads one key per feature group. Its Rust fetcher runs inside your model service and used about 1 ms of CPU per request in a production deployment.
- Chalk merges every overlapping bucket in its Query Servers on each request.
- Databricks allows at most four instances per online store and 4,096 concurrent requests per workspace.
- Snowflake sizes its online store for hundreds to low thousands of queries per second, with at most 10 entities per query.
- Feast scales as far as the store and pipelines you build.
5–9x cheaper to serve at 100,000 requests per second
At high request rates the bill is mostly online serving, and Zipline has no serving tier to pay for.
| Monthly cost on AWS | 10k req/s | 100k req/s | 1M req/s |
|---|---|---|---|
| Zipline | $13.9k | $48.4k | $484k |
| Chronon (open source) | $119k8.5x | $164k3.4x | $666k1.4x |
| Chalk | $43.7k3.1x | $260k5.4x | $2.6M5.4x |
| Databricks | $45.5k3.3x | $443k9.2x | $4.4M9.2x |
| Snowflake | $85.5k6.1x | $753k15.6x | $7.4M15.4x |
Peak requests per second, 20 million entities, 20 feature groups per request (5 streaming), high availability everywhere, list prices. Zipline includes its usage-based fee (1x infrastructure); open-source Chronon includes four engineers to run it. Ratios are against Zipline.
Chalk runs every request through Query Servers that its docs say to keep under 50% CPU, and bills them at $0.18 per vCPU-hour. Databricks bills Model Serving by provisioned concurrency, and merging sawtooth rows makes each request slower. Snowflake needs about 800 autoscale nodes at 100,000 requests per second. Open-source Chronon has no GigaTile, so it stores and reads recent tiles for every streaming feature, runs a separate fetcher service, and needs people to run it. Zipline's total includes its usage-based fee.
Zipline picks the online store from your traffic: Redis for IO-heavy workloads, DynamoDB on AWS or Bigtable on GCP for storage-heavy ones. The calculator applies the same choice to every platform that runs on a key-value store, and lets Zipline put batch and streaming groups in different stores.
Training and serving compute the same thing
Zipline runs one aggregation library for backfills, streaming and serving, and checks online values against offline ones for you.
| Where training and serving can differ | Built-in consistency check | |
|---|---|---|
| Zipline | Backfill, streaming and serving share one library | Consistency, drift, coverage |
| Chronon (open source) | Same shared library | A job you run |
| Chalk | The oldest online bucket counts in full, while offline windows are exact | No |
| Databricks | Python UDFs return None offline and NaN online; streaming Last ignores event order online | No |
| Snowflake | Offline reads depend on the session time zone, a bug Snowflake has acknowledged | No |
| Feast | Batch and streaming pipelines are separate code | No |
Sources
Prices are list prices from November 2026, in AWS us-east-1 and GCP us-central1: Databricks Model Serving and limits, Lakebase, Databricks Feature Views, Snowflake credit consumption and online feature store, Chalk pricing, Query Server sizing and materialized aggregations, ElastiCache, DynamoDB, Memorystore for Valkey, Bigtable and the Feast docs. How GigaTile works is in Approaches to Streaming Features.
Serving cost calculator
Hover a bar for its line items, or open the full list below.
Assumptions the vendors don’t publish
Show line items
| Platform | Line item | Monthly |
|---|---|---|
| Zipline | ElastiCache Valkey· 15 shards, 1 replica each | $7.7k |
| Cross-zone reads | $4.4k | |
| Flink streaming· 188 vCPU | $5.5k | |
| Spark batch (spot) | $6.7k | |
| Zipline usage-based fee· 1x infrastructure | $24.2k | |
| Total | $48.4k | |
| Chronon (open source) | ElastiCache Valkey· 67 shards, 1 replica each | $34.2k |
| Cross-zone reads | $7.7k | |
| Fetcher service· 1,020 vCPU | $27.0k | |
| Flink streaming· 188 vCPU | $5.5k | |
| Spark batch (spot) | $6.7k | |
| Engineers to run it· 4 FTE | $83.3k | |
| Total | $164k | |
| Chalk | ElastiCache Valkey· 38 shards, 1 replica each | $19.4k |
| Cross-zone reads | $13.1k | |
| Query Servers· 1,250 vCPU at 50% CPU | $167k | |
| Streaming· 188 vCPU | $25.0k | |
| Background persistence· 53 vCPU | $7.1k | |
| Batch and backfill | $28.5k | |
| Total | $260k | |
| Databricks | Model Serving endpoints· 6,000 concurrency | $307k |
| Lakebase online store· 267 x CU_8 in 67 stores | $108k | |
| Streaming pipelines | $18.0k | |
| Batch materialization | $10.5k | |
| Total | $443k | |
| Snowflake | Online service· M + 800 autoscale nodes | $703k |
| Batch refresh (warehouse) | $48.0k | |
| Online sync (warehouse) | $2.2k | |
| Online storage | $93 | |
| Total | $753k |