SynapCores v2.0 — Run your transactions, your analytics and your ML on one engine
Stop paying for four systems to answer one question. v2.0 makes SynapCores a
database you can put money through, point a BI tool at, and train models in —
without moving a byte between vendors.
Put money through it: transactions that survive the power going out
The capability: financial-grade correctness for the workloads you currently
keep in a separate OLTP database.
If your platform moves money, allocates inventory, or books capacity, "probably
committed" is not a state you can reconcile in the morning. v2.0 gives you
crash-durable transactions: when a COMMIT returns success, that data has
reached the write-ahead log and will survive a hard power loss. Other
connections cannot read your half-finished work, and SAVEPOINT lets a
multi-step operation back out one stage without abandoning the whole thing.
What this retires: the reconciliation job that exists because you didn't
fully trust the last write. And the second database you were running because
this one couldn't be trusted with the ledger.
Proof: we kill the server with SIGKILL mid-transaction and restart it —
zero partial rows. We acknowledge a 100-row commit, kill it, restart —
all 100 rows present, with every constraint still enforced. 16 of 16
crash and concurrency gates pass on the shipped build.
Available on Community and Enterprise. No configuration required — it is the default.
Read the durability guarantees →
Stop bad data at the door: referential integrity that is actually enforced
The capability: your schema's own rules now do the work your application
code was doing.
Declare a FOREIGN KEY and v2.0 enforces it — orphan rows are refused at
insert, cascades run inside the same statement, and ON DELETE SET NULL does
what it says. Table-level UNIQUE is enforced too, single-column and composite.
Why this matters commercially: every integrity rule that lives in
application code is a rule each new service has to re-implement, and one of them
eventually won't. Moving those rules into the database makes them true for every
client, including the ones your team hasn't written yet.
Before you upgrade, run one query. Earlier versions accepted these
declarations and never checked them, so you may already hold rows that violate
constraints you believe are active. v2.0 refuses new violations but will not
retroactively clean existing data — the audit query is in the upgrade guide.
We would rather you find this on your terms than in production.
Available on Community and Enterprise.
Get the pre-upgrade audit query →
Answer analytics questions on your own storage, 2–3× faster than DuckDB
The capability: query Parquet in your own S3 bucket, in place, at speeds
that make interactive exploration realistic.
| What you're doing |
SynapCores v2.0 |
DuckDB 1.5.5 |
|
| Scan 189,361 rows out of a lake table |
10.1 ms |
28.6 ms |
2.8× faster |
| Count a join across two lake tables |
5.5 ms |
10.9 ms |
2.0× faster |
| Join returning 176,400 rows |
16.5 ms |
29.1 ms |
1.8× faster |
GROUP BY one text column |
5.0 ms |
3.8 ms |
1.3× slower |
Same Parquet objects, same bucket, measured server-side on both engines, with
every answer compared row by row. We publish the row we lose, because a
benchmark that only shows wins tells you nothing about the one query you
actually care about. If your workload is dominated by wide GROUP BY
aggregation, DuckDB is still the faster reader for that shape.
What this unlocks: dashboards that aggregate lake data in milliseconds
instead of seconds, and exploratory joins fast enough to actually explore with.
No warehouse load step, no copy, no second vendor holding your data.
Available on Community and Enterprise.
See the benchmark harness →
Move production MySQL into your own S3, on a schedule, with proof it arrived
The capability: a managed export pipeline that replaces an ETL tool, a
scheduler and the glue between them.
Point SynapCores at a production MySQL table, give it your bucket, and it writes
Parquet to object storage on a cron you set — then registers that Parquet as an
ordinary SQL table you can query in place.
The operational risk it removes: the 3am run that silently half-worked.
Every run reports rows written, bytes, peak memory and the child process exit
code, and verified_partitions tells you every file the job claimed to write
was read back and confirmed. Credentials are encrypted at rest and the API
never returns your secret — only the key id, so you can audit which credential
is in use without being able to leak it.
Who this is for: teams running warehouse queries against a transactional
database, where the fix has been "buy a warehouse" and the real problem is that
nobody wants to own another pipeline.
Available on Community and Enterprise. Requires an admin to configure the destination credential.
Follow the 20-minute walkthrough →
Train, forecast and segment without moving data to a notebook
The capability: 32 machine-learning algorithms, trained with SQL, against
the data already in your database.
v2.0 adds 19 algorithm paths, including two capabilities that were simply absent:
Find your customer segments without labelling anything. Clustering —
KMeans, DBSCAN, Gaussian mixtures and more — needs no target column at all.
Hand it behavioural columns and it finds the groups that are actually in your
data, instead of the thresholds somebody guessed in a spreadsheet three years
ago.
Forecast demand straight from SQL. Seasonal-naive, ETS, ARIMA, SARIMA and an
additive changepoint model, with chronological validation — because a random
train/test split on a time series quietly leaks the future and flatters the
score.
Keep models current without owning an MLOps stack. A deployed model can
retrain itself on a schedule that survives restarts, or when its inputs drift
away from what it was trained on. Drift detection needs no labels — it
compares live traffic against the distribution the model was fitted on — so it
works on production data where the outcome isn't known yet. The industry answer
to this is a monitoring library plus a separate orchestrator; here it is one
clause in a SQL statement.
And money columns finally reach the model. A DECIMAL price or balance used
to arrive as a missing value and get replaced by the column average — your
model could not see spend at all. Fixed.
Honest scope: these are native implementations in the spirit of the
well-known libraries, not those runtimes embedded, and they do not read or
write upstream model files. Clustering quality is reported as a separation
score, not an accuracy, because there is no ground truth to be accurate
against. We would rather tell you that than let you present a silhouette as a
78% hit rate.
Available on Community and Enterprise. Optuna-based tuning requires a Python runtime on the host.
Train your first model in SQL →
Know when your database is actually unhealthy
The capability: a readiness probe that reflects whether SQL works, not just
whether a process is alive.
A liveness check that only proves the HTTP listener is up will happily report
green while queries hang — so nothing restarts, and nobody is paged. v2.0 adds a
readiness endpoint that genuinely exercises the SQL path, and makes shutdown
cooperative: on SIGTERM the engine stops accepting connections, drains the
ones in flight, finishes accepted training work and exits cleanly, so a rolling
deploy stops looking like a crash.
For platform teams: point livenessProbe at the cheap check and
readinessProbe at the real one, and your orchestrator will finally take a
wedged instance out of rotation.
Available on Community and Enterprise. Requires admin access to configure probes.
Get the Kubernetes probe configuration →
The trade-off we chose, and the number
Single-row writes are about 2× slower than v1.18. Roughly +0.7 ms on an
individual INSERT, UPDATE or DELETE by primary key.
That is not a regression we missed — it is the cost of the durability guarantee
at the top of this page. Every write now reaches the write-ahead log before it
is acknowledged, which is precisely what makes an acknowledged commit survive a
power cut. You cannot have the first property without paying for it.
Reads did not regress — every read case in our 39-case guard was within
noise or faster, several lake shapes materially so.
If you are write-throughput-bound and don't need crash atomicity, one
environment variable restores the previous write path. You give up isolation,
crash atomicity and foreign-key enforcement to do it, which we think is the
wrong trade for most teams — but it is your call, and it is documented.
Batching amortises it: wrap many rows in one transaction and the per-statement
cost largely disappears.
Upgrading
Pull the new image or binary and restart. No schema migration, no client changes.
Three things to check first:
- Audit for constraint violations you already hold —
FOREIGN KEY and
table-level UNIQUE were not enforced before v2.0.
- Add retry handling for concurrent writers — real isolation means real
conflicts. Catch the retryable error and retry the whole transaction. Code
that never saw a conflict may start seeing them; that is the isolation
working.
- If you send a bare
BEGIN to the stateless query endpoint, it now fails
loudly instead of silently autocommitting — which is the one change that can
break working-looking code, though what that code was doing was never a
transaction.
Read the full upgrade guide →
Known issues
- Inside an explicit transaction, writes to a table carrying a trigger or a
durable-agent binding are refused rather than silently excluded from the
commit. Outside a transaction they work normally. A trigger's side effect is
not part of the same commit as its row — full atomicity covers your rows and
constraints, not hook effects.
AUTOML.PREDICT inside a common table expression requires the feature columns
to appear in that CTE's own projection. The workaround is one line and is in
the docs.
- Two algorithm names are accepted by the parser for compatibility but have no
implementation; asking for them is refused immediately with a list of what is
available, rather than failing after you have waited.
Every number on this page was measured on the build this release ships, on a
non-AVX-512 consumer CPU — not a tuned benchmark rig. Validation evidence,
including the comparisons where we lose, ships in the repository.