Repository navigation
Release HyperDX - #3720
Open
github-actions[bot] wants to merge 2 commits into
Open
Release HyperDX#3720github-actions[bot] wants to merge 2 commits into
github-actions[bot] wants to merge 2 commits into
Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
github-actions
Bot
force-pushed
the
changeset-release/main
branch
8 times, most recently
from
November 19, 2026 15:43
b75f3a7 to
6bee557
Compare
github-actions
Bot
force-pushed
the
changeset-release/main
branch
17 times, most recently
from
November 20, 2026 20:34
d451153 to
ba9bbf0
Compare
github-actions
Bot
force-pushed
the
changeset-release/main
branch
from
November 20, 2026 21:10
d7c16b7 to
a1b13c4
Compare
This branch was successfully deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.
Releases
@hyperdx/cli@0.9.2
Patch Changes
event_patterns,emerging_signalsand search denoise, and the CLI,no longer split the same log line by its date. Before mining, the Drain miner
masks the date of a timestamp (weekday, day, month, year) to
<DATE>, so datedlines from different days, months or years get the same template instead of one
"new" and one "gone". Times and ids are not masked. Within one day lines group
as before, except that klog lines with different severity letters are no longer
merged. No mask changes a line's token count: the
event_patternsdrill-downpredicate matches
<DATE>as one or more non-whitespace characters, also insidea token (the new
masksoption of the pattern predicate builders, which onlythe MCP drill-down sets), and search denoise masks bodies the same way before
matching. Masking instructions get an optional
perWordflag (default off, asin drain3). The web app's Event patterns view uses its own drain3 config and is
unchanged.
@hyperdx/api@2.47.0
Minor Changes
Patch Changes
22377df: perf: bound the check-alerts previous-history lookup by the alert interval
The check-alerts task no longer scans 7 days of alert history per alert on
every tick. A one-row lookup per alert skips alerts already evaluated in the
current window. Non-grouped alerts take their previous state from that row,
and grouped alerts scan a window sized to their interval (1 to 7 days).
b129ec1: feat: Classify failed ClickHouse proxy requests on the server span and the
clickhouse_proxy.querySLI (error_category), including errors ClickHousewrites after a 200 response. The span also records the upstream status, the
ClickHouse error code and type, and the body of upstream errors that carry no
ClickHouse code.
86acf27: perf: read only 20 median-sized parts for text index key and value lookups
Map key discovery and "all values" lookups on text indexes read the index
dictionary of every part in the time range, which is slow on tables with many
parts. They now read the 20 parts closest to the median size. Log and trace
sources can change the count with the new "Map key index parts" setting
(
mapKeyIndexMaxParts), and which parts are read with "Map key index part size"(
mapKeyIndexPartSelection: median, smallest, or largest). "Load more" on asingle filter key still reads every part in the time range.
e80fc1b: fix(mcp): explain hasToken() separator errors in query tools
When a SQL filter calls
hasToken()with a needle that contains_,-,.,@,:or spaces, ClickHouse fails with "Needle must not contain whitespace orseparator characters". MCP query tools now append a hint pointing to a Lucene
term, or in SQL to
position()for the exact term (withhasAllTokens()as anoptional text-index speed-up), and the
whereLanguagedescription warns aboutit.
2f61137: fix: return partial rows from
clickstack_searchinstead of timing outThe MCP search tool now scans newest-first over the same progressive time
windows as the app's search page, under a 25s budget. When time runs out it
returns the rows found so far with a
partialblock (requested range, the rangefully covered, and the window that was cut off) and a hint for the next call,
instead of failing after 30s. The time-window helpers moved from the app to
@hyperdx/common-utils/core/searchWindows; the app's behavior is unchanged.64a55d5: fix: allow grouping gauge and sum metric charts by an expression over the time column
A table, pie or bar chart on a gauge or sum metric grouped by a time expression
such as
toStartOfHour(TimeUnix)ortoDate(TimeUnix)failed withUnknown identifier TimeUnix, because the bucketed CTE the outer query readsdid not carry the time column. When the group-by references it,
TimeUnixnowresolves to each metric bucket's start, so values are the same as the chart
without the time group-by.
e8d1d9a: fix: return the actual earliest/latest time for min/max on a date or time column
minandmaxon a Date, DateTime, or DateTime64 column (for exampleTimestamp) returned 0, because every aggregation cast its value to Float64first. Tables, number tiles, and the MCP
clickstack_tabletool now return thetimestamp itself. Line, bar, pie, and other charts, ratios, formulas, and number
tiles with several series keep the numeric cast.
d99be76: fix: MCP
event_patterns,emerging_signalsand search denoise, and the CLI,no longer split the same log line by its date. Before mining, the Drain miner
masks the date of a timestamp (weekday, day, month, year) to
<DATE>, so datedlines from different days, months or years get the same template instead of one
"new" and one "gone". Times and ids are not masked. Within one day lines group
as before, except that klog lines with different severity letters are no longer
merged. No mask changes a line's token count: the
event_patternsdrill-downpredicate matches
<DATE>as one or more non-whitespace characters, also insidea token (the new
masksoption of the pattern predicate builders, which onlythe MCP drill-down sets), and search denoise masks bodies the same way before
matching. Masking instructions get an optional
perWordflag (default off, asin drain3). The web app's Event patterns view uses its own drain3 config and is
unchanged.
e58329c: fix: require the password when changing a connection's host
Updating a connection's host without sending its password used to keep the stored password, so it would be sent to the new host on the next query. Both
PUT /connections/:idandPUT /api/v2/connections/:idnow return 400 in that case. Prometheus endpoints never use the stored password, so it is cleared instead. The connection form asks for the password when the host is edited and shows the API's error message.2cec94f: feat: let SQL query dashboard filters use an optional source
A SQL query filter can now pick a source on its connection, or a data source variable, and can do so on a connection variable too. With one, its SQL can use
$__sourceTableand$__filters, which applies the dashboard's filter selections for that source, the same way raw SQL tiles do. The external API and MCP read and write it assourceId.7afa31c: feat: Add a team-wide default (All or Exact) for the search filters "Show All Values" setting
Updated dependencies [e36cec3]
Updated dependencies [1a4c116]
Updated dependencies [351f194]
Updated dependencies [86acf27]
Updated dependencies [2f61137]
Updated dependencies [64a55d5]
Updated dependencies [e8d1d9a]
Updated dependencies [d99be76]
Updated dependencies [2cec94f]
Updated dependencies [7afa31c]
Updated dependencies [f11b9f5]
@hyperdx/app@2.47.0
Minor Changes
Patch Changes
6e7ee83: feat: Removing the NEXT_PUBLIC_ENABLE_ADHOC_FILTERS flag
1a4c116: perf: load search filter values faster in exact mode by reading several keys on
one Map column with a single
groupUniqArrayMap, and by querying plain columnsand each Map or JSON column separately so slow attribute maps don't hold up the
rest
351f194: feat: allow up to 10,000 in the team "Filter Keys Fetch Limit" setting
The setting's input in Team settings stopped at 1,000 keys per column; it now
accepts up to 10,000. Map key discovery returns up to the setting's value per
Map column, and never fewer than 1,000.
0c77d52: fix: stop the previous filter type's validation from blocking a save after switching types
Switching a dashboard filter from "PromQL label values" to another type no longer fails with "Select a PromQL source".
661a5a1: fix: address JSON array elements by index in row-panel filters
"Add to Filters" on a value inside an array in a native JSON column now uses
1-based array subscripts, for example
events[1], instead of quoted keys suchas
events.0, so the filter matches rows instead of silently returning none.The same applies to arrays inside a JSON string (for example a String
Bodycolumn), which now pass 1-based integer indexes to
JSONExtract*. JSON stringsstored in a native JSON column are wrapped in
toString()soJSONExtract*accepts them.
3d12509: fix: keep JSON-column filter keys in JSON form when another filter is added
Adding a second filter, or clicking "Load more" on a filter, no longer rewrites
an earlier filter on a key inside a native JSON column to
col['key'], whichClickHouse rejects (
Illegal type Dynamic of argument of function in, orFirst argument for function 'arrayElement' must be array, got 'JSON'on olderversions).
86acf27: perf: read only 20 median-sized parts for text index key and value lookups
Map key discovery and "all values" lookups on text indexes read the index
dictionary of every part in the time range, which is slow on tables with many
parts. They now read the 20 parts closest to the median size. Log and trace
sources can change the count with the new "Map key index parts" setting
(
mapKeyIndexMaxParts), and which parts are read with "Map key index part size"(
mapKeyIndexPartSelection: median, smallest, or largest). "Load more" on asingle filter key still reads every part in the time range.
2f61137: fix: return partial rows from
clickstack_searchinstead of timing outThe MCP search tool now scans newest-first over the same progressive time
windows as the app's search page, under a 25s budget. When time runs out it
returns the rows found so far with a
partialblock (requested range, the rangefully covered, and the window that was cut off) and a hint for the next call,
instead of failing after 30s. The time-window helpers moved from the app to
@hyperdx/common-utils/core/searchWindows; the app's behavior is unchanged.64a55d5: fix: allow grouping gauge and sum metric charts by an expression over the time column
A table, pie or bar chart on a gauge or sum metric grouped by a time expression
such as
toStartOfHour(TimeUnix)ortoDate(TimeUnix)failed withUnknown identifier TimeUnix, because the bucketed CTE the outer query readsdid not carry the time column. When the group-by references it,
TimeUnixnowresolves to each metric bucket's start, so values are the same as the chart
without the time group-by.
e8d1d9a: fix: return the actual earliest/latest time for min/max on a date or time column
minandmaxon a Date, DateTime, or DateTime64 column (for exampleTimestamp) returned 0, because every aggregation cast its value to Float64first. Tables, number tiles, and the MCP
clickstack_tabletool now return thetimestamp itself. Line, bar, pie, and other charts, ratios, formulas, and number
tiles with several series keep the numeric cast.
0c77d52: feat: rename the "Queried values" dashboard filter type to "SQL expression values"
The new name separates it from "SQL query values" and "PromQL label values".
e58329c: fix: require the password when changing a connection's host
Updating a connection's host without sending its password used to keep the stored password, so it would be sent to the new host on the next query. Both
PUT /connections/:idandPUT /api/v2/connections/:idnow return 400 in that case. Prometheus endpoints never use the stored password, so it is cleared instead. The connection form asks for the password when the host is edited and shows the API's error message.a0732b1: feat: save a temporary dashboard with its content intact
67e2eb3: Use the sidebar layout icons for showing and hiding search filters
2cec94f: feat: let SQL query dashboard filters use an optional source
A SQL query filter can now pick a source on its connection, or a data source variable, and can do so on a connection variable too. With one, its SQL can use
$__sourceTableand$__filters, which applies the dashboard's filter selections for that source, the same way raw SQL tiles do. The external API and MCP read and write it assourceId.0c77d52: feat: autocomplete table and column names in the SQL query filter editor
The SQL query dashboard filter's editor now suggests the tables and columns of
the selected connection's sources, like the Raw SQL tile editor.
7afa31c: feat: Add a team-wide default (All or Exact) for the search filters "Show All Values" setting
6a5fd11: In dark mode, the field background now has opacity so it works on different
backgrounds, such as popovers. It is set directly in the theme tokens instead of
through a Click UI token that resolved to a solid color.
Updated dependencies [22377df]
Updated dependencies [b129ec1]
Updated dependencies [e36cec3]
Updated dependencies [1a4c116]
Updated dependencies [351f194]
Updated dependencies [86acf27]
Updated dependencies [e80fc1b]
Updated dependencies [2f61137]
Updated dependencies [64a55d5]
Updated dependencies [e8d1d9a]
Updated dependencies [d99be76]
Updated dependencies [e58329c]
Updated dependencies [2cec94f]
Updated dependencies [7afa31c]
Updated dependencies [f11b9f5]
Updated dependencies [7905721]
@hyperdx/common-utils@0.36.1
Patch Changes
e36cec3: perf: share one Map key read between concurrent callers
The search page's filter panel, search bar and group-by each load the same Map
keys when the page opens. Each sent its own text index read, which on large
tables can use several GiB of memory. Concurrent requests for the same keys now
share one query.
1a4c116: perf: load search filter values faster in exact mode by reading several keys on
one Map column with a single
groupUniqArrayMap, and by querying plain columnsand each Map or JSON column separately so slow attribute maps don't hold up the
rest
351f194: feat: allow up to 10,000 in the team "Filter Keys Fetch Limit" setting
The setting's input in Team settings stopped at 1,000 keys per column; it now
accepts up to 10,000. Map key discovery returns up to the setting's value per
Map column, and never fewer than 1,000.
86acf27: perf: read only 20 median-sized parts for text index key and value lookups
Map key discovery and "all values" lookups on text indexes read the index
dictionary of every part in the time range, which is slow on tables with many
parts. They now read the 20 parts closest to the median size. Log and trace
sources can change the count with the new "Map key index parts" setting
(
mapKeyIndexMaxParts), and which parts are read with "Map key index part size"(
mapKeyIndexPartSelection: median, smallest, or largest). "Load more" on asingle filter key still reads every part in the time range.
2f61137: fix: return partial rows from
clickstack_searchinstead of timing outThe MCP search tool now scans newest-first over the same progressive time
windows as the app's search page, under a 25s budget. When time runs out it
returns the rows found so far with a
partialblock (requested range, the rangefully covered, and the window that was cut off) and a hint for the next call,
instead of failing after 30s. The time-window helpers moved from the app to
@hyperdx/common-utils/core/searchWindows; the app's behavior is unchanged.64a55d5: fix: allow grouping gauge and sum metric charts by an expression over the time column
A table, pie or bar chart on a gauge or sum metric grouped by a time expression
such as
toStartOfHour(TimeUnix)ortoDate(TimeUnix)failed withUnknown identifier TimeUnix, because the bucketed CTE the outer query readsdid not carry the time column. When the group-by references it,
TimeUnixnowresolves to each metric bucket's start, so values are the same as the chart
without the time group-by.
e8d1d9a: fix: return the actual earliest/latest time for min/max on a date or time column
minandmaxon a Date, DateTime, or DateTime64 column (for exampleTimestamp) returned 0, because every aggregation cast its value to Float64first. Tables, number tiles, and the MCP
clickstack_tabletool now return thetimestamp itself. Line, bar, pie, and other charts, ratios, formulas, and number
tiles with several series keep the numeric cast.
d99be76: fix: MCP
event_patterns,emerging_signalsand search denoise, and the CLI,no longer split the same log line by its date. Before mining, the Drain miner
masks the date of a timestamp (weekday, day, month, year) to
<DATE>, so datedlines from different days, months or years get the same template instead of one
"new" and one "gone". Times and ids are not masked. Within one day lines group
as before, except that klog lines with different severity letters are no longer
merged. No mask changes a line's token count: the
event_patternsdrill-downpredicate matches
<DATE>as one or more non-whitespace characters, also insidea token (the new
masksoption of the pattern predicate builders, which onlythe MCP drill-down sets), and search denoise masks bodies the same way before
matching. Masking instructions get an optional
perWordflag (default off, asin drain3). The web app's Event patterns view uses its own drain3 config and is
unchanged.
2cec94f: feat: let SQL query dashboard filters use an optional source
A SQL query filter can now pick a source on its connection, or a data source variable, and can do so on a connection variable too. With one, its SQL can use
$__sourceTableand$__filters, which applies the dashboard's filter selections for that source, the same way raw SQL tiles do. The external API and MCP read and write it assourceId.7afa31c: feat: Add a team-wide default (All or Exact) for the search filters "Show All Values" setting
f11b9f5: perf: read large-block Map text index values per key, most frequent first
"All values" lookups on Map key=value text indexes with large dictionary blocks
(2048 or more, e.g.
text_index_dictionary_block_size = 8192, which thecollector's migration now sets on ClickHouse 26.6+)
now read each key separately, in parallel queries of at most 10 keys, and stop
after 20,000 tokens per key, so a key with a value per row (request IDs, UUIDs)
can't hold up the others. Values come from
topKWeighted(limit)weighted byeach token's row count, so a capped list holds the most common values, most
common first. These reads no longer apply
max_rows_to_read: the index is readin sorted token order, so the cap left every key that sorts after a
high-cardinality key with no values.
Indexes with smaller blocks, including ClickHouse's default of 512, keep the
previous query shape (one read for all of a query's keys, the alphabetically
first values, and the
max_rows_to_readcap), now in parallel queries of atmost 10 keys, so the cap's cut-off affects at most 9 other keys. Every read loads
each part's sparse index, which is large with small blocks, so one read per key
would multiply that cost. The block size comes from the index definition, then
the table's
text_index_dictionary_block_size, then the server's.Text indexes on plain columns now return the most frequent values with
topKWeighted, in parallel queries of at most 10 columns, without the row cap.@hyperdx/otel-collector@2.47.0
Patch Changes
d945d7e: fix(security): bump
clickhouse-go/v2to v2.51.0 in the migration toolPulls in
github.com/andybalholm/brotliv1.2.7, which fixes GO-2026-6888(CVE-2026-111497), a denial of service in the brotli decompressor.
f11b9f5: perf: use 8,192-token text index dictionary blocks on the seeded tables
On ClickHouse 26.6 and later, the collector's schema migration now sets
text_index_dictionary_block_size = 8192(the default is 512) onotel_logs,otel_tracesandhyperdx_sessions, withALTER TABLE … MODIFY SETTINGafterthe seed. Each part's text index sparse index is 16x smaller. Reading filter
values with
mergeTreeTextIndex()loads that index for every part on everyread, and on a 537M-row table 4,096-token blocks already made those reads about
10x cheaper per part, made searches for a rare attribute value 3-7x faster, and
made searches for a common value 10-30% faster.
Because it's an
ALTER, existing installs get it too: new parts use 8,192-tokenblocks, and existing parts keep theirs until they're merged. A table that
already sets
text_index_dictionary_block_sizeis left alone, and operators whomanage these tables themselves can turn the change off with
HYPERDX_OTEL_EXPORTER_SET_TEXT_INDEX_BLOCK_SIZE=false. ClickHouse 26.2–26.5doesn't have the setting, so those servers keep 512-token blocks.