Problemsby GlyphClickHouse upgrades break working setups
31 posts from 20 people19d activeFading, +0 this week
Posts per day
The posts behind it
Iceberg mutations write position delete files that reads apply to many data files; one file per data file cuts `apply_patch_parts_merge` reads by 89–91%
### Describe the situation Iceberg `ALTER TABLE ... UPDATE/DELETE` writes one position delete file per partition with the deleted rows of every data file it touches, so later reads apply each such file to many data files and read it for each. Writing one delete file per data...
DEFAULT of a column added by ALTER is evaluated in the global context for rows written before the ALTER (MergeTree, Memory)
### Describe the issue For rows written before `ALTER TABLE ... ADD COLUMN c ... DEFAULT `, the default is evaluated at read time in a copy of the global context, not in the context of the reading query. MergeTree does this for a part without the column...
`_server_credentials_allowed` changes the identity of a `disk(...)` declared before 26.7, so it can't be re-declared after an upgrade
...rejects. There seems to be no SQL workaround: - The injector strips any marker the user writes before it decides whether to add one ([DiskFromAST.cpp#L214](https://github.com/ClickHouse/ClickHouse/blob/v26.8.34.13-lts/src/Disks/DiskFromAST.cpp#L214)). - Without the session...
Failed ALTER RENAME COLUMN rollback on MergeTree leaves the rename mutation applied against mismatched metadata
Two holes in the rollback of a failed `ALTER … RENAME COLUMN` on plain `MergeTree` (`StorageMergeTree::alter`, the `catch` after `addPreparedMutationEntry` / `setProperties`). Both leave the table in a state where a merge applies the rename mutation against metadata that does...
A `CREATE` that fails to publish the table leaves the inner tables of a `MaterializedView` or `TimeSeries` table
### Company or project name _No response_ ### Describe what's wrong A `CREATE` of a `MaterializedView` with an inner table, or of a `TimeSeries` table, that fails while the table is being published leaves the inner tables it has already created in the database, without an owner....
DNSCacheUpdater spins forever (100% CPU, holds DNSResolver mutex) in `CaresPTRResolver::wait_and_process` after the DNS server port becomes unreachable; `SYSTEM DROP DNS CACHE` and shutdown hang
### Describe what's wrong When the background `DNSCacheUpdater` refreshes a cached **reverse** (PTR) entry and the DNS server's UDP port becomes closed (the server answers with ICMP port unreachable, e.g. while the DNS service restarts), the updater thread never leaves...
Growing a large PODArray is 1.4–2× slower and needs ~1.5× more memory since mmap/mremap left Allocator (#52792)
**Company or project name** **Describe the situation** Since 23.9 (#52792, first attempt #51686 reverted in #52589) `Allocator::realloc` goes through jemalloc for every size. A large buffer that grows by doubling (`PODArray::push_back` / `insert`, `HashTable::resize`,...
Wide parts with a `Geometry` column written before 26.8 are unreadable after the upgrade: `Stream g.MultiPoint.size0 ... is not found`
...the part and other variant subcolumns (`g.Point`, `g.LineString`) works. There is no SQL workaround on 26.8: `ALTER ... MODIFY COLUMN` and `OPTIMIZE` need to read the column. Rolling back to 26.6 makes the data readable again, but a part merged on 26.6 again has no `MultiPoint`...
Answer repeated count() with a filter from exact per-granule counts in the query condition cache
### Describe the situation A repeated `SELECT count() FROM t WHERE ` still reads every granule that has a match, because the query condition cache keeps one bit per granule. Keeping the exact number of matching rows per granule lets `count()` add them up instead of reading....
ClickHouse: Drill to detail and SQL Lab table preview fail on AggregateFunction columns
### Bug description ### Bug description On ClickHouse tables with `AggregateFunction(...)` columns (the usual AggregatingMergeTree rollup), Drill to detail, dataset samples and the SQL Lab table preview fail with: These features select every column as-is, so the query returns...
Iceberg INSERT doesn't cut data files at `iceberg_insert_max_bytes_in_data_file` when Parquet encodes in parallel (regression from #127670)
Since #127670, `MultipleFileWriter` starts a new data file when `buffer->count()`, the number of bytes already written to the current file, reaches `max_data_file_num_bytes`...
ppc64le: bundled libunwind unw_getcontext leaves callee-saved vector registers byte-swapped (fixed in LLVM main, not backported)
On little-endian ppc64le, `unw_getcontext` in contrib/llvm-project/libunwind/src/UnwindRegistersSave.S corrupts the caller's callee-saved vector registers. PPC64_STVS(n) does `xxswapd n,n; stxvd2x n,0,4` and never swaps back, so v20-v31 and f14-f31 come back byte-swapped. Fixed...
RFC: Compression codecs for subcolumns
...`ZSTD` for a string); today one `CODEC` clause applies to every leaf, and the only workaround is to split the column. ### Syntax In `CREATE TABLE`, a keyword-first entry: the keyword `CODEC`, the name of the target, and the codec chain in parentheses. Like `INDEX`, `CONSTRAINT`,...
`url` `ArrowStream` reads fail on HEAD 501: new seekability probe rethrows
...= 0` avoids it. - Any server with chunked/close-delimited GET and a failing HEAD breaks every `url`/`URL`-engine `ArrowStream` query. Root cause: `ArrowIPCBlockInputFormat::readStreamBatch` now calls `MessageReader::readBody` which unconditionally calls `checkIfActuallySeekable`...
`comparisonCanThrow` skips `Variant` inside `Tuple`: reordered `WHERE` throws
...values holding `Variant`/`Dynamic` (for example behind a JSON/string predicate) starts failing with `NO_COMMON_TYPE` after the upgrade; workaround `short_circuit_function_evaluation_reorder_arguments = 0`. The same misclassification reaches the `force_enable` lazy-argument...
Heavy argument shared by several `and` is now evaluated per `and`: 2.9x slower
### Describe what's wrong - A heavy predicate shared as the first argument of several `and`/`or` calls runs once per call instead of once overall. - Triggered when `short_circuit_function_evaluation_reorder_arguments` is 1 (the default) and several `countIf` calls share the same...
Accepted `disk(...)` definition rejected on RESTORE and re-ATTACH
...and (by code, not run) a new `Replicated` database replica cannot create it. Loud error; workaround is ` 1 ` in the server config, which also disables the server-settings check. ### Does it reproduce on most recent release? No - only `master` (`a5b019f44f7b`). The bug starts...
Lens worker claims job but never progresses, fails after ~15 min with "Worker disconnected repeatedly"
...surface a clear, actionable error instead of silently retrying for 15 minutes and then failing with a generic message that gives no indication of *why* the worker disconnected. **Verified**: the `ValidationError` warning in the worker log; every run failing with `"Worker...
`additional_result_filter` runs over the extremes stream: `LOGICAL_ERROR`
### Describe what's wrong - With `extremes = 1`, `additional_result_filter` evaluates the predicate on the synthetic min/max extremes rows as if they were data. - Trigger: any `SELECT` with `extremes = 1` and `additional_result_filter`, for example filter `a + b >= 2` over...
Update the zstd fork and evaluate the pending upstream decompression work
### Why `ZSTD(3)` is now the default compression codec of MergeTree columns, marks and primary keys, so zstd decompression is the largest part of many scans. For example, in `SELECT COUNT(DISTINCT SearchPhrase) FROM hits` (ClickBench Q5) with the radix partitioned `uniqExact` of...
Parquet v3 reader: a negative `num_values` in a column-chunk footer trips `chassert(diff.by_stage[i] >= 0)` in `ReadManager`
🕵️ Reading a Parquet file whose footer declares a negative `ColumnMetaData.num_values` for a column chunk aborts `clickhouse` on debug and sanitizer builds. The abort is a `chassert` in the v3 reader's memory accounting: reached from `ReadManager::flushMemoryUsageDiff`...
Background column-TTL merge wedges forever when a dependent MATERIALIZED column throws on the reset default (regression from #117815)
A background column-TTL merge permanently wedges when a dependent `MATERIALIZED` column's expression throws on the source column's default value (e.g. `intDiv(100, x)` when the TTL resets `x` to `0`). Introduced on master by #117815: the recompute added there runs inside the TTL...
Pending async inserts into `Distributed` become permanently undeliverable after `RENAME COLUMN` and wedge all later sends
**TL;DR:** With default settings, an `INSERT` into a `Distributed` table is acknowledged and queued for background delivery. A later, perfectly legal `ALTER TABLE ... RENAME COLUMN` (applied to the local and the `Distributed` table) makes every pending file permanently...
Compact codec proof reads only the first frame: `NONE`-default part refused
...because the first-frame check in the Compact leg fails to detect the default codec. Actual output of the reproducer above on `master` (`4127c65654d0`): Suggested fix In the Compact leg, walk the frames of `data.bin` with `getCompressionCodecForFile(..., skip_to_next_block =...
hasAny over Array(String) compares every element with every constant needle: a needle hash set cuts a 32-needle hasAny from 5.4 s to 0.52 s
### Describe the situation `hasAny(arr, [constants])` over an `Array(String)` column compares every element with every needle through a virtual `IColumn::compareAt`, because `sliceHasImplAnyAll` is specialised only for numeric arrays. A row without a match therefore costs...
[Bug]: langfuse_otel builds OTLP endpoint without required /v1/traces suffix — self-hosted Langfuse traces silently dropped (404)
...`/api/public/otel` suffix (as opposed to the generic `/v1/traces` default) - there is no way to fix this via env vars today for `langfuse_otel` short of overriding the entire endpoint including the correct suffix, which the current code does not support (the `/api/public/otel`...
Text index prunes matching rows: decline keys on needle padding, not value's
### Describe what's wrong - `WHERE s = toFixedString(...)` on a `String` column silently prunes rows whose value equals the needle plus trailing zero bytes. - Triggered by `text(tokenizer = array)` or `text(tokenizer = splitByString([sep]))` indexes when `use_skip_indexes = 1`,...
Reuse checksum-matching detached parts when bootstrapping a fresh replica
### Use case When creating a fresh ReplicatedMergeTree replica, we want to seed it from an object-storage backup instead of transferring the entire dataset from peers. Backup parts can be restored into the table’s detached/ directory. Attaching them normally may allocate new...
BACKUP/RESTORE ... TABLE db.t AS t loses the rename through parseQueryToJSON / formatQueryFromJSON (RESTORE target silently changes database)
`BACKUP`/`RESTORE ... TABLE db.t AS t` loses the rename through `parseQueryToJSON` / `formatQueryFromJSON`. For `RESTORE` this changes what the query does: the `AS` clause is the destination, so `RESTORE TABLE d.t AS t FROM ...` (restore into the current database) silently...
ALTER NAMED COLLECTION DELETE lacks dependency check, breaks table reload
...engine definition needs ` ` from `nc`. - The table keeps working after the ALTER but fails to load on next ATTACH or server restart with a required-keys error. - `DROP NAMED COLLECTION` correctly refuses this same case with NAMED_COLLECTION_IS_USED, but ALTER DELETE does not. -...
Implement compact snapshot files for `plain_rewritable` disks to speed up initialization and updates.
### Company or project name ClickHouse ### Describe the situation Initialization of a `plain_rewritable` disks can take a lot of time, as it has to list the metadata directory, read all `prefix.path` files, and list all the related directories. It can be tens of thousands of S3...
Companies and products named
About this problem
Evidence
31 posts from 20 people in 3 places, about 4 a week over 45 days. Mostly on ClickHouse/ClickHouse, BerriAI/litellm, apache/superset. Tools named alongside: GitHub, Text, Langfuse, Kubernetes.
Frustration Frustration 0 of 3· Seen on GitHub issues
Summary, model-written
ClickHouse users face performance issues and unexpected behavior with array comparisons, default ZSTD compression, table codecs, and Iceberg delete files.
How it was grouped
Posts that state a pain and match the "breaking-changes" rule about ClickHouse. First post Sep 21, 2026, latest Nov 4, 2026. Corroborated: 3 or more posts from 2 or more people or places. Method
The same problem elsewhere
GitHub upgrades break working setups
89 posts from 77 people on GitHub issues, Hacker News, Stack Overflow.
Upgrades and new releases break working setups
77 posts from 63 people on GitHub issues, Reddit, App Store reviews.
Claude Code upgrades break working setups
68 posts from 61 people on GitHub issues, Hacker News, Reddit.
Codex upgrades break working setups
67 posts from 60 people on GitHub issues, Reddit.
DuckDB upgrades break working setups
45 posts from 35 people on GitHub issues.
Claude upgrades break working setups
35 posts from 34 people on GitHub issues, Reddit, Hacker News.
Anthropic upgrades break working setups
27 posts from 22 people on GitHub issues.
PyTorch upgrades break working setups
26 posts from 21 people on GitHub issues, Reddit.
ChatGPT upgrades break working setups
26 posts from 26 people on GitHub issues, Community forums, App Store reviews.
OpenAI upgrades break working setups
25 posts from 25 people on GitHub issues, Reddit, Community forums.
Open WebUI upgrades break working setups
25 posts from 2 people on GitHub issues.
Gemini upgrades break working setups
24 posts from 24 people on GitHub issues, Community forums, Reddit.
Other ClickHouse problems
ClickHouse users keep filing bugs
Top post: "Tighten seccomp policy for optional kernel interfaces", one of 19 posts from 8 people on GitHub issues.
Trouble with ClickHouse column
Top post: "MaterializedPostgreSQL: reading a Nullable column's .null subcolumn (x IS NULL, count(x)) fails with", one of 15 posts from 9 people on GitHub issues.
ClickHouse costs outgrow the value it gives
11 posts from 3 people on GitHub issues.
Trouble with ClickHouse table
Top post: "`SYSTEM RESTORE REPLICA` on a second replica over-bumps the table `metadata` znode version; every later", one of 8 posts from 5 people on GitHub issues.
Experiments on ClickHouse are hard to reproduce and track
8 posts from 1 person on GitHub issues. Unverified.
Trouble with ClickHouse tables
Top post: "Paimon: resolve columns by field id, so tables with renamed or re-added columns are readable", one of 7 posts from 6 people on GitHub issues.
Trouble with ClickHouse join
Top post: "Explicit grace_hash now spills at max_bytes_ratio_before_external_join = 0.5 although the build side fits in", one of 5 posts from 4 people on GitHub issues.
ClickHouse error messages are cryptic
5 posts from 3 people on GitHub issues.
Trouble with ClickHouse read
Top post: "ORC: a corrupt PostScript `metadata_length` makes ORC predicate pushdown read `stripe_stats` out of bounds,", one of 4 posts from 4 people on GitHub issues.
Trouble with ClickHouse database
Top post: "`DROP NAMED COLLECTION` succeeds while a database uses the collection, and the server then fails to start", one of 4 posts from 2 people on GitHub issues.
Trouble with ClickHouse query
Top post: "Memory tracker counts jemalloc's freed pages as used, so a heavy query started right after another spills to", one of 4 posts from 3 people on GitHub issues.
ClickHouse rate limits and usage caps block work
3 posts from 3 people on GitHub issues.
History