Repository navigation
v0.10.1
Search by meaning — class matching and MITRE technique matching — could stop finding classes and
techniques written after the database restarted, and fell back to name and type matches without a
word. Deployments now repair their vector indexes themselves: class indexes are checked on every
platform start, after a module reset and at the 10-minute vector-search check; MITRE indexes on the
first MITRE match after start and at the same check. Search by meaning works again after the upgrade.
The bundle now ships Memgraph 3.13.2; moving an existing deployment to it is optional. Compared
against the previous tag, v0.10.0.
Upgrading: take the new bundle and follow the operator guide's upgrade procedure — back up,
unpack, set PLATFORM_VERSION, ./byodt update. That procedure changes only PLATFORM_VERSION: an
existing deployment keeps the DB_IMAGE in its .env, and the only other line of .env.example that
changed is DB_IMAGE. The repair works on Memgraph 3.8.1 as it is, so a deployment that stays on
3.8.1 is repaired by this release alone, and bumping DB_IMAGE is optional. To move to Memgraph
3.13.2, back up data/memgraph first — there is no downgrade — then set
DB_IMAGE=docker.io/memgraph/memgraph:3.13.2 in .env in the same change as PLATFORM_VERSION,
so the first start on 3.13.2 is also the first start on 0.10.1 and repairs any index carried over
from 3.8.1. ./byodt up and ./byodt update now warn when .env's DB_IMAGE differs from the
image the bundle was tested with; the warning never blocks. The reference-data ingest is skipped,
since the corpus did not change. The general module's content hash was restamped for a corrected
load-balancer rule (see Fixed), so the platform's load-time skip gate reinstalls it; an existing
finding picks up the correction the next time its element's attributes or class binding are written.
On the first start the platform logs Vector index … rebuilt: holds N of N embedded … classes for
each class index it rebuilds, and the first MITRE technique match logs MITRE vector index … rebuilt
for each MITRE index; a deployment whose indexes are already correct logs nothing. No saved recipe
needs regenerating.
Changed
- The bundle ships Memgraph 3.13.2.
.env.exampleandNOTICEname
memgraph/memgraph:3.13.2(still under the Business Source License 1.1). Unlike the 3.8 releases,
3.13.2 does not re-key a vector index when it recovers from a snapshot. Existing deployments keep
theirDB_IMAGE(see Upgrading). byodtsays when the database is not the tested one.byodt1.1.0'supandupdateprint a
warning, with the backup-first steps to move, when.env'sDB_IMAGEdiffers from the image the
bundle was tested with. The comparison is on the image'sname:tagonly, so a reference to the same
image through a registry mirror does not warn. They never block.- A module whose metadata cannot be read at boot still registers its API. A module whose
getMetadata()throws — for example because its backing service is unreachable — now registers its
schema and resolvers under its directory name, with one warning and no retries, and is flagged
metadataUnavailable. Its classes are installed on a later boot that reads the metadata. Modules
withoutgetMetadata()or without a name are still rejected.
Fixed
- Search by meaning finds every class and technique again. Memgraph 3.8.x can recover a vector
index from a snapshot under a different label or property than it was created on; writes then miss
the index, so classes installed, and MITRE nodes written, after such a restart were missing from
class matching and MITRE technique matching. The platform now checks the class vector indexes on
every start — even when no module changed — and after a single-module reset, and checks the MITRE
vector indexes on the first MITRE match after start. An index keyed on another label or property,
or holding fewer entries than there are embedded nodes, is dropped and created again. While the
platform runs, the first match after the 10-minute vector-search check expires repeats the check,
so a restart of the database alone is repaired without restarting the platform. A snapshot written
while an index was mis-keyed keeps the wrong key on a later Memgraph too, so the repair is needed
on any database version. - Deleting a class or a MITRE node no longer breaks search by meaning on the newer database. On
Memgraph 3.13.2 a node deleted after it was indexed is still returned by vector search until the
database's garbage collection runs — every five minutes with the bundle's settings — and reading it
failed the match. Every vector hit is now matched again by its internal id, which skips deleted
nodes, in class matching, MITRE technique matching and the class index check. Memgraph 3.8.1
already skipped such nodes. - A class match restricted to modules finds those modules' classes. With
moduleIds, class
matching searched three timestopNnearest neighbours and then applied the module filter, so it
returned nothing whenever other modules' classes were nearer. It now searches every embedded class
of the label before filtering. - A control can be created without a control class or a folder. The data-access library sent an
empty class link and an empty folder link with such a create, and the API rejects a link input
with nothing to connect; it now leaves them out. The plugin bundles the library, so
@dether.net/dethereal0.4.11 carries the fix. - A failed request is reported as an error, not as missing data. With
errorPolicyallor
ignore, the data-access library returned no data and callers failed with "Cannot read properties
of undefined".performQueryandperformMutationnow throw aDtRequestErrorthat carries the
server's message alone, capped at 1000 characters. GraphQL errors and 4xx responses are not
retried; 5xx responses and transport failures are, as before. Partial data with errors passes
through, and clients onerrorPolicynoneare unaffected. - The load balancer's weak TLS termination rule fires. One arm of the general module's
load-balancer "weak / legacy TLS termination" rule comparedmin_tls_versionwith values outside
its enum, so it never fired; it now reads the enum's weak value. The module content hash is
restamped (see Upgrading).
Documentation
- The backend references describe vector index health. The class and control resolver
specification, the custom resolver services reference, the module management service reference and
the backend delegation design describe the re-keyed index, when the platform verifies and rebuilds
class and MITRE indexes, the module-scoped search limit, and the re-match by internal id that skips
deleted nodes. The MITRE vector tier reference also
drops HNSW parameters the platform does not set, and states that one unhealthy MITRE kind no longer
turns off the others. - The upgrade guide no longer moves the database with every release. The operator guide's
upgrade section now describes the database move as optional, one-way and backup-first, instead of
asking to take everyDB_IMAGEchange together with the release.