sqlglot is the right default. It is excellent, and it is MIT-licensed — which means you may ship it inside your own commercial product for free, and we are not going to pretend otherwise. It handles query-level lineage on cloud warehouses well enough that DataHub reports 97–99% accuracy on it. If it covers your SQL, use it. General SQL Parser is what teams buy at the two points where free runs out: when the SQL that matters is procedural, and when shipping a parser to your own customers means you need a counterparty rather than a warranty disclaimer.
Stay on sqlglot if you are in Python, your SQL is queries, and your dialects are in its set. Buy GSP for one of two reasons, not for a feature list: because your transformation logic lives in stored procedures and dynamic SQL that a tolerant parser returns as opaque nodes, or because you are redistributing a parser inside your own product and need a contract rather than a warranty disclaimer. Most teams who should buy GSP keep sqlglot in place and escalate to it.
This is the part that decides most commercial evaluations, and it has nothing to do with grammar quality. MIT lets you ship sqlglot in your product — we are not going to argue otherwise. It also disclaims every warranty, which means the risk stays with you.
| sqlglot (MIT) | GSP (commercial) | |
|---|---|---|
| Shipping it inside your commercial product | Permitted. MIT lets you redistribute — with all warranty expressly disclaimed and no indemnity to you or your customer. | A Distribution License: a named contract covering redistribution in your product, with Gudu carrying obligations rather than disclaiming them. |
| When a customer’s SQL breaks the parser | You open a GitHub issue. Maintainers are generous but owe you nothing, and the fix lands when it lands. | You send the failing statement and it is scheduled into a release. That is the obligation you are buying, more than the grammar. |
| Version commitment | You self-manage. Pinning an old version means you stop receiving fixes for it. | You choose the upgrade window; a pinned release keeps receiving fixes under maintenance. |
| Procurement, security and legal review | No counterparty. There is nobody to sign a DPA, complete a security questionnaire, or answer your customer’s auditor. | A Hong Kong company on the other side of a contract: security review, VPAT, and audit questions get answered by a person. |
Where sqlglot is better, this table says so. A comparison that has the vendor winning every row is not worth reading.
| Capability | sqlglot | GSP |
|---|---|---|
| Language / runtime | Python (optional Rust tokenizer) | Java (JDK 8+) and .NET — in-process, no Python runtime |
| License | MIT open source | Commercial, with a 90-day free trial |
| Dialect-to-dialect transpilation | Yes — the headline feature (e.g. Snowflake → DuckDB), and better at it than GSP | No. GSP parses, validates, formats, rewrites and analyses; it is not a transpiler |
| Dialects | 30+ dialects, one shared grammar with per-dialect overrides | 42 dialects, each with a dedicated grammar — including IBM Db2, Informix, Netezza, Sybase, SAP HANA, Vertica, N1QL, SOQL, GaussDB, Dameng, OceanBase and Power Query M (15 of them on .NET) |
| Query-level lineage on cloud warehouses | Strong. DataHub reports 97–99% accuracy for query lineage built on sqlglot — for SELECT/CTAS/INSERT on Snowflake, BigQuery and Databricks this is the correct default | Equivalent for this workload. Buying GSP for it alone is not a good reason |
| Stored procedures / procedural SQL | Procedural blocks generally fall back to opaque Command nodes; production T-SQL stored procedures remain a known open gap | Fully parsed ASTs: PL/SQL packages, types and triggers; T-SQL blocks, TRY/CATCH, EXECUTE; Db2 SQL PL cursors; Teradata procedures and BTEQ |
| Dynamic SQL | Not evaluated | Evaluated — GSP constant-folds EXEC(’…’) strings and parameter bindings to recover lineage from dynamic SQL |
| Multi-statement scripts | Not a supported unit of analysis | Parsed as a script, with lineage carried across statements |
| Syntax validation | Best-effort — a tolerant parser by design | Offline validation with error positions, no database connection |
| SQL formatting | A pretty-print flag | Dedicated formatter, 72+ style options |
Mostly you should not. If your SQL is queries against cloud warehouses and you are in Python, sqlglot is the right answer and we would use it too. There are two situations where people buy GSP instead: the SQL that matters to them is procedural (stored procedures, packages, triggers, dynamic SQL) and comes back as an opaque node, or they need to ship a parser inside a commercial product and want a counterparty rather than a disclaimer. If neither is true for you, keep your money.
Yes, you can, and we will not pretend otherwise — MIT permits redistribution. What it does not do is transfer any risk. The licence disclaims all warranties, so when your customer hits a parse failure in production, the cost and the deadline are yours, and there is no one contractually obliged to help. A Distribution License is not about permission; it is about who carries that obligation. For some products that is worth nothing, and for products sold into regulated enterprises it is often the whole reason for the purchase.
Some of it goes away, and we would rather say so than be caught out. sqlglot is well resourced — its parent Tobiko Data was acquired by Fivetran in September 2025 — and procedural support has been on their roadmap. The parsing-depth gap should be expected to narrow over time. What does not narrow is the contractual side: redistribution with indemnity, a scheduled fix path, and an accountable counterparty are structurally unavailable from an open-source project no matter how good the code becomes. If you are choosing on grammar depth alone, re-test both against your own SQL every year.
Only by embedding a Python runtime or standing up a sidecar service. Both work and neither is exotic, but we are not aware of a named .NET or JVM product vendor doing it in production — the operational cost usually outweighs the licence saving. GSP ships native Java and .NET editions that run in-process. Note that on the JVM you do have free in-process alternatives (JSqlParser, Apache Calcite); on .NET the free options are T-SQL-only or syntax-only, which is a materially different situation.
That is the arrangement we recommend most often. Keep sqlglot as the default path for transpilation and query-level work, and escalate to GSP only for what it returns an error or an opaque node on — typically stored procedures, dynamic SQL, and the legacy dialects. You pay for the exceptions rather than replacing something that already works.