Axapta 3.0 and AX 4.0 store batch and cleanup times as a local date plus
seconds. They were turned into UTC with the SQL Server's offset at collection
time, so every time from the other side of a DST change was an hour off (a job
planned 01:50 CET on the spring night looked 70 minutes late at 03:00 CEST),
and a SQL Server running UTC behind a CET AOS shifted every time.
AxLocalTime converts with the rules of the AX time zone on that date:
SQLInsight:DynamicsAx:LocalTimeZone (Windows or IANA id), else the agent's own
zone when its offset matches the SQL Server's, else the server's offset (to the
nearest quarter hour) with a Danish data note. The skipped spring hour and the
repeated autumn hour get standard time (deterministic, never throws, never a
false "late"); seconds outside 0..86400 are unknown and 86400 is midnight. The
AX 3.0 cleanup time is local by version, also when LATESTCLEANTIME is missing.
AX 2009+ UTC times are untouched. Also: remaining numbers cannot overflow for
HIGHEST = Int64.MaxValue, and run keys use the invariant culture.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The AX scripts written in 1.4 (own header style and a generator) now follow
main's "SQL Scripts" structure: the review header, "-- Kilde:" blocks
registered in SqlScriptSources and checked word for word against the code.
- SqlScriptSources: the shared AX detection (DatabasesSql, LoginSql,
BuildDetectionSql), companies, batch jobs, batch history (first read and
from the watermark) and number sequences / RecId for Axapta 3.0, AX 4.0,
AX 2009 and AX 2012; BlockingCollector split 2008 / 2005, as CONTEXT_INFO
comes from sys.dm_exec_sessions from 2008 and from sysprocesses on 2005.
- SQLI-AX-001..010 and _collectors/Ax*.sql in main's format; COM-AX-001..007
with the 1.4 queries, the AX detection and USE [AXDB]; BLK-001,
BlockingCollector and DatabaseDiscovery updated (CONTEXT_INFO,
sp_executesql).
- Scripts_are_read_only stays strict (no INTO #temp anywhere) and now also
checks the batch inside EXEC [db].sys.sp_executesql N'...'.
- SQLINSIGHT_UPDATE_SQL_SCRIPTS=1 (from 1.4) rewrites out-of-date blocks from
the code; the 1.4 generator test now checks that every AX script shows its
sources for every AX version and says what is unverified.
- README: SQLI-AX-001..010, the AX collector scripts, 108 rules.
- AxDatabaseDetector: LoginSql and BuildDetectionSql as members, so the
registry sees them.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The detector looks again every 5 minutes at an AX database where the agent
lacks part of the read access, so a GRANT takes effect quickly. The AX SQL
test checks that only standard AX tables are read. The hub end-to-end test
expects no AX rule outcomes without the opt-in. Docs: what 1.5's synthetic
AX checks can reuse.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
0017_dynamics_ax adds insight.AxBatchRuns (completed batch runs, EventDays
retention) and insight.AxSequenceSamples (one NEXTREC/NEXTVAL sample per
sequence per day, HourlyDays retention). The agent reads them back after a
restart so job comparisons and forecasts continue, takes its thresholds from
SQLInsight:DynamicsAx, and only sends findings of the built-in SQLI-AX rules
to the hub for databases with the AX opt-in ticked; AX rule reasons (which
can name tables, logins and GRANT text) are replaced by a fixed text.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
dynamics-ax (hourly) lists the monitored AX databases with version, schema,
companies and read access; discovery and the AX community rules (now marked
dynamicsAx) use the same detection. dynamics-ax-batch (5 min) reads BATCH
(3.0/4.0) or BATCHJOB/BATCHJOBHISTORY (2009+), keeps each job's own run
history; dynamics-ax-sequences (hourly) reads NUMBERSEQUENCETABLE and
SYSTEMSEQUENCES and forecasts from SQLInsight's own daily samples. All reads
run through [db].sys.sp_executesql under read committed with a 2 s lock
timeout (snapshot when allowed), bounded with TOP, never NOLOCK, never
written. Blocking now captures CONTEXT_INFO for the AX user mapping.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The 2012+ query used SELECT ... INTO #system_health, a write to tempdb.
One SELECT with ROW_NUMBER() now returns the deadlock rows and the newest
row (the read position) in the same pass. SQL Scripts and the read-only
test no longer allow any exception.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Registers 1.3's SQL in SqlScriptSources (DatabaseDiscovery: databases 2005/2000
and the Dynamics AX signature per database; PermissionCheck: login/sysadmin, the
HAS_PERMS_BY_NAME checks for 2005-2012, 2014-2019/MI and 2022+, databases without
access) and adds _collectors/DatabaseDiscovery.sql and PermissionCheck.sql with
Danish review headers. The constants are renamed *Sql and the login query moved to
PermissionCheck.WhoSql so the registry test finds them. The scripts of the
collectors that now skip unchosen databases say so (their SQL is unchanged).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
CollectorContext carries the instance's DatabaseSelection; the database,
index, Query Store and community-rule collectors skip databases that are not
chosen (system databases always run). DatabaseDiscovery lists every database
with state, size, recovery model, compatibility level and access, and
recognises Dynamics AX with the same signature as the AX rules, through
three-part names only (2000 and 2005+ variants). PermissionCheck reports which
rights from grant-least-privilege.sql the login has (HAS_PERMS_BY_NAME,
version-aware; not possible on 2000).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A DBA should be able to review every statement SQLInsight sends to a
monitored server before granting it access. "SQL Scripts/<rule-id>.sql"
holds, per rule (82 built-in, 16 community), a Danish header (what and
why, how C# judges the data, thresholds with config keys, severity,
exact GRANTs, versions, collector file and member) and the collector's
SQL verbatim, every version variant as its own batch.
"_collectors/<Collector>.sql" holds each collector's queries once.
The only statement that is not strictly read-only is documented in the
scripts: the deadlock collector's session temp table (#system_health).
Inline SQL in the collectors moved to internal constants and Build*Sql
methods (no text changed), so SqlScriptTests can check that the scripts
still contain the SQL the code sends, that every rule has a script, that
every collector query is registered, and that nothing writes.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
With ALTER TRACE the error log collector read every default trace file
(up to 5 x 20 MB) every 5 minutes. While the current file is the one read
last time, everything since then is in it, so only that file is read;
after a rollover (and the first time) all files are read once as before.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
sys.dm_tran_top_version_generators reads the whole version store. It ran
every 5 minutes while the version store was over half the threshold,
i.e. exactly when it is large and expensive to scan, and a timeout failed
the whole tempdb collection. It now runs at most once per hour (the
previous answer is kept in between, with its time) and a failed scan
keeps the previous answer.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
When the distribution database could not be read, SQL Server's message
(which names the agent's login) was appended to Unavailable. That text is
the rules' "not measurable" reason and is returned unredacted by
/rule-states and /replication. It is now UnavailableDetail: shown on the
local instance page and [udeladt] in the API unless
ApiIncludesSensitiveEvidence. The agent list is also capped (TOP 500).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The open transactions and long requests had no TOP, so a busy server could
return (and store in the snapshot every minute) thousands of rows with
statement texts. Now at most 100 transactions (large log first, then the
oldest) and the 50 longest requests.
The transactions were joined to sys.dm_exec_requests, which has one row
per request: a MARS session with an open transaction gave duplicate rows,
and the MERGE into insight.LongTransactions then failed on the primary
key every minute. The first request is now taken with OUTER APPLY TOP 1,
and the save keeps one row per transaction.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The default trace was filtered on whole seconds only, so the newest entry
read (ReadUntil) came back at every collection until a newer one was
written: one 824 became a new event every 5 minutes. Entries at or before
ReadUntil are now left out, as for xp_readerrorlog.
The ring buffer's times were computed to whole seconds from the clock at
each read and moved by up to a second between reads, so the newest entry
could be counted twice. They are now computed to the millisecond with a
small tolerance.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The availability group collector also reads sys.dm_hadr_cluster_members
(empty or unavailable without a WSFC), and SQLI-AG-003 (High) reports a
node or witness that is not up, with how many members are still up.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The index collector now sums index reads and writes between collections
across restarts of SQL Server (and of the agent, by seeding from the
stored snapshot) and keeps them per day in insight.IndexUsageDaily
(migration 0015, HourlyDays). SQLI-IDX-002 judges an index once it has
been followed for 7 days instead of needing 7 days of uptime.
GET /api/v1/instances/{name}/index-usage lists usage per index.
SQLI-STAT-001 also reports statistics that were never updated on tables
with rows; statistic rows now come from the table's heap or clustered
index, as a never-updated statistic has no row count of its own.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
On a distributor, a 15-minute collector reads each distribution agent's
latest run (status, latency, message) and replication errors the last
day from the distribution database. That needs SELECT on
MSdistribution_agents, MSdistribution_history and MSrepl_errors, which
the agent does not have by default (never the replicated commands);
without it SQLI-REPL-001 (agent failed/retrying, High) and SQLI-REPL-002
(latency > 5 min, Medium) are "not measurable" with the exact grant, and
"not relevant" elsewhere. Shown on the instance page and in
GET /api/v1/instances/{name}/replication. Tested against a real
distributor set up in the SQL 2022 container.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A five-minute collector reads tempdb space (free, version store, user and
internal objects, whether files can grow), the 'Version Store Size (KB)'
counter, the sessions using most tempdb, the oldest snapshot transaction
and, only when the version store is large, the databases generating
versions. Trend and heavy sessions survive restarts; used space and the
version store become metrics with explanation and assessment.
Rules SQLI-TEMP-001 (version store > 1 GB or steadily growing),
SQLI-TEMP-002 (nearly full and cannot grow) and SQLI-TEMP-003 (session
above 1 GB for 15 minutes), configurable under SQLInsight:Tempdb.
"tempdb lige nu" on the instance page; GET /api/v1/instances/{name}/tempdb.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A one-minute collector reads open user transactions (age, log bytes used
and reserved, database, session, login, program, statement) and user
requests running longer than the threshold, leaving out the agent's own
and system sessions, BACKUP/RESTORE/DBCC (long operations) and requests
waiting on purpose. Transactions open a minute or holding 100 MB are kept
in insight.LongTransactions (migration 0014, RawDays).
Rules SQLI-TRAN-001 (open > 10 min, High at 6x), SQLI-TRAN-002 (> 1 GB
log) and SQLI-REQ-001 (statement > 15 min), configurable under
SQLInsight:Activity. Shown "right now" on the instance page and in the
time window analysis and PDF; GET /api/v1/instances/{name}/activity and
the analysis JSON leave logins and statements out.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A five-minute collector reads important error log entries from the best
source the agent's rights allow: the error log itself (xp_readerrorlog,
only with securityadmin), the default trace's ErrorLog and Audit Login
Failed events (optional ALTER TRACE), or the exception and connectivity
ring buffers (VIEW SERVER STATE: numbers and severity, failed logins).
Repetitions are summed per error; the watermark is the newest entry
read, because the default trace is buffered. Stored in
insight.ErrorLogEvents (migration 0013, EventDays), resumed after restart.
Rules SQLI-LOG-001..006: corruption 823/824/825, I/O over 15 s (833),
severe errors (17+), login failure bursts (configurable), stack dumps and
AG errors; rules needing the log text are "not measurable" with the
right to grant. Page /instanser/{navn}/fejllog (menu Overvågning) and
GET /api/v1/instances/{name}/error-log with messages redacted.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- Long operations: a one-minute collector reads operations with progress
from sys.dm_exec_requests (BACKUP, RESTORE, DBCC, REORGANIZE, SHRINK,
ROLLBACK): command, database, start, percent, expected end, login.
Stored in insight.LongOperations (migration 0011, RawDays), shown on
the instance page, in the time window analysis and its PDF. The last
10 durations per command and database survive restarts; SQLI-OPS-001
flags a backup/restore running over twice the median.
- Full-text community rules COM-FT-001..005 (metadata only, 2005 syntax):
disk-full or paused/throttled catalogs, stale manual/off change
tracking indexes and disabled indexes.
- Statement activity (CPU time / time in plan cache): column in top SQL
and SQLI-CPU-002 when above SQLInsight:TopSql:ActivityThresholdPercent
(default 50) for at least an hour. Never clears the plan cache.
- BuiltInCollectors.Create takes a CollectorSettings record.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A new light collector samples sys.dm_exec_requests every 15 seconds:
waiting user sessions (not the agent's own) by wait type, database and
statement (same key as top SQL), with blocking. Sampling-only cycles
reuse the last probe and skip rule evaluation. Stored in
insight.ActiveWaitSamples (migration 0010) for RawDays.
Wait types on the instance page and in the analysis link to
/instanser/{navn}/ventetyper/{type}: Danish explanation from a new
knowledge base (shared with the SQLI-WAIT rules, with Dynamics AX notes),
trend over time, the statements that waited with share and estimated
wait, a link to the full text (/instanser/{navn}/sql) and, for lock
waits, to blocking in the same window. SQL Server 2005 compatible.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Collects members of the fixed server roles and CONTROL SERVER every 15
minutes with login type, disabled state and SQL login password policy.
Windows groups are unfolded recursively through Active Directory on a
Windows agent (cached, 20 s timeout, failures shown with the reason).
Changes since the previous collection are logged, and who/when is read
from the default trace (event 108) when the agent has the optional
ALTER TRACE; authors that only reach the buffered trace later are filled
in at the next collection. The first collection is the baseline.
New rules SQLI-SEC-001 (new privilege, High for sysadmin/securityadmin/
CONTROL SERVER) and SQLI-SEC-002 (more sysadmins than the limit, default
3). Names are sensitive evidence and never in titles or subjects.
Web: /instanser/{navn}/rettigheder and /rettigheder with change log by
period, alert banner on the instance page and overview, and an audit
PDF (/rapporter/rettigheder.pdf) linked from Rapporter.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- new collector instance-config: collation, edition, version, services and
accounts (sys.dm_server_services, 2008 R2 SP1+), Instant File
Initialization and Lock Pages in Memory (where the build has the
columns), tempdb files, default paths; plus a fingerprint of the whole
setup (sp_configure, trace flags, services, version, tempdb files)
- changes are computed against the previous collection, which the agent
restores from the repository after a restart; migration 0008 keeps them
in insight.ConfigurationChanges (as long as resolved findings)
- SQLI-CFG-013 "Konfiguration ændret": information for a day when
sp_configure, a trace flag or a service account changes (account names
are sensitive evidence)
- page /instanser/{navn}/konfiguration: instance, CPU and memory, services,
tempdb, trace flags, non-default sp_configure (default vs. in use vs.
configured) and the change log; linked from the instance page
- the management report lists configuration changes in the period
- trace flag reading shared by trace-flags and instance-config
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- file-io records each file's drive: sys.dm_os_volume_stats from 2008 R2 SP1
(new capability VolumeStats, mount points included), otherwise derived
from the path (D:\..., \server\share\, / on Linux), so 2005 works too
- migration 0007: Volume and IntervalSeconds on FileIoSamples, and
DiskHourly (rolled up by maintenance, kept HourlyDays)
- DiskActivity sums files per drive: read/write IOPS, MB/s and latency
weighted by I/O count
- instance page: "Diskhastighed pr. drev" (table and MB/s + latency charts);
time window analysis and its PDF: per drive against the same hours of the
four previous weeks
- SQLI-IO-001/002 now report per drive (one finding when the storage is
slow, the slowest files in the evidence) instead of per file
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
At a customer whose repository sits on a monitored instance, the agent
reported 12 deadlocks that were its own: two instances' monitors save
findings in parallel, one transaction each, and while insight.Findings is
small SQL Server scans it, so the two lock each other's rows.
- SaveRuleResultsAsync takes an exclusive application lock for its
transaction (the agent is the only writer; a save takes milliseconds)
and retries if chosen as deadlock victim anyway
- repository connections are named "SQLInsight Pro (repository)"
- deadlocks where every participant is SQLInsight Pro are not reported
(the read position still moves past them); any other session keeps it
The new integration test (two instances saving 25 rules x 40 rounds in
parallel) failed 2 of 2 without the fix and passed 5 of 5 with it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Seen on a customer's SQL Server 2005 SP4 (legacy ODBC):
- agent-jobs declared @since with an initial value (2008+): "Incorrect syntax
near CONVERT" and SQLI-JOB-001/002 never ran. Now DECLARE; SET.
- plan-analysis used query_hash (2008+) and failed every 30 minutes for
SQLI-PLAN-001/002/003. On 2005 a statement is now identified by
sql_handle + offsets, as top SQL does. It needs sys.dm_exec_text_query_plan,
so the new capability TextQueryPlan (2005 SP2, 9.0.3042) is required.
- COM-DB-001 used OBJECT_SCHEMA_NAME (2005 SP2+); now joins sys.objects and
sys.schemas.
Audit of all collector SQL and community rules without minimumVersion 10.0:
the rest is 2005 syntax or behind a version/capability check.
Sql2005CompatibilityTests (2005 cannot run in CI): every collector that can
run on 2005 is scanned for 2008+ syntax in its SQL (inline DECLARE, +=,
MERGE, CONCAT/IIF/FORMAT/TRY_*, OFFSET/FETCH, date/time types, multi-row
VALUES), version-branched files listed explicitly; the same for community
rules; and the 2005 variants of plan analysis and top SQL must not use
query_hash. Verified to fail on the old agent-jobs SQL.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
New collector query-stats reads sys.dm_exec_query_stats for the statements
that ran since the previous collection, grouped by query_hash (SQL Server
2008+) or sql_handle + offsets (2005, via legacy ODBC; literals only). Only
the 50 heaviest by CPU, duration, reads and executions, text cut to 4000
characters, no plan XML. The database comes from the plan's dbid attribute,
because ad hoc batches (as Dynamics AX sends) have none in sql_text.
QueryStatsDelta diffs against the previous collection: a statement seen
before with the same plan gets the difference, a plan compiled since then
counts in full, a recompiled or evicted plan or counters that went
backwards start over (never negative), a restart gives no delta. Up to 2000
statements are remembered between collections.
Repository migration 0006: insight.QueryStatSamples (deltas, top 100 per
collection) and insight.QueryTexts (text once per statement), kept for
Retention:RawDays. The text is customer data: never stored as a JSON
snapshot and never part of hub submissions. Window queries for metrics,
waits (raw and hourly), file I/O, blocking, deadlocks, findings active in a
window and top SQL by CPU/duration/reads/executions.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Rules can now be data: a JSON file with one read-only SELECT where every
row is a finding, with {column} templates, server or per-database scope,
minimum version, sensitive columns and references. Community rules live
in rules/community (reviewed via pull request), ship with every build
and with the signed rule pack; local rules are read from the agent's
data folder every five minutes. Queries are checked for anything that
could write, run with the least-privilege login, and a missing
permission makes the rule not measurable rather than failing. The
rule engine picks up catalog changes without a restart. Authors get
'rules validate' and 'rules test --instance'. Four example community
rules (sa active, orphaned database owners, default trace off,
untrusted foreign keys).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E3EExwndy8Z22EK2Qpt6BK
- Managed Instance reports 12.0.2000.8 but is evergreen: treated as the
newest engine with all capabilities; new SelfManagedInfrastructure
capability so backup, CHECKDB, files, tempdb, memory, patch level and
lifecycle rules report 'handled by Azure' instead of false findings
- Microsoft Entra ID login (managed identity, service principal, default
credential chain) via SqlConnection.AccessTokenCallback
- Least-privilege script recognizes Managed Instance; deadlock collector
falls back to the ring buffer when event files cannot be read
- Web UI shows the right release for Managed Instance
- README and installation docs: notifications and Managed Instance
- Catalog test: 61 rules with unique, well-formed ids
Integration suite green on SQL Server 2017 and 2022.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E3EExwndy8Z22EK2Qpt6BK
- Availability group collector (replica and database replica states,
queues, commit times) and rules: SQLI-AG-001 unhealthy/disconnected
replica or suspended data movement (scripts HADR RESUME), SQLI-AG-002
secondary behind (estimated data loss and redo time)
- SQL Agent collector using only granted msdb tables (no EXECUTE on msdb
functions): SQLI-JOB-001 failed job with the failing step's message
(backup/CHECKDB jobs are High), SQLI-JOB-002 SQL Agent not running
- Integration tests create a real availability group (CLUSTER_TYPE NONE)
and a failing job in the test container (HADR and Agent enabled)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E3EExwndy8Z22EK2Qpt6BK
Collector (every 6 h, metadata and DMVs only, no scans): index inventory
with usage and forwarded fetches, missing-index suggestions, statistics
modifications via sysindexes.rowmodctr (needs only metadata visibility).
Rules:
- SQLI-IDX-001 missing index, evaluated against existing indexes:
extends an index with the same keys instead of adding one; notes on
write-heavy tables, many indexes, wide includes, short uptime
- SQLI-IDX-002 unused index (needs 7 days of uptime; suggests DISABLE)
- SQLI-IDX-003 duplicate or covered index (filters, uniqueness respected)
- SQLI-IDX-004 heap with forwarded records
- SQLI-STAT-001 stale statistics, SQLI-STAT-002 auto statistics off
Column lists use U+E000 as separator: FOR XML cannot serialize NCHAR(1).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E3EExwndy8Z22EK2Qpt6BK
Collector takes the 50 heaviest statements (CPU, last 24 h), fetches
statement plans as text (dm_exec_query_plan returns NULL for deeply
nested plans) and analyzes them in C#. Rules:
- SQLI-PLAN-001 implicit conversion preventing an index seek
- SQLI-PLAN-002 join without predicate (deliberate CROSS JOIN excluded)
- SQLI-PLAN-003 non-inlined scalar UDF in a heavy statement
- SQLI-PLAN-004 spills to tempdb, SQLI-PLAN-005 oversized memory grants
(2016 SP2+ / 2017 CU3+ columns, new capability)
Verified against real plans in SQL Server.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E3EExwndy8Z22EK2Qpt6BK
Collector reads the 100 heaviest queries per Query Store database, split
into the last 24 hours and the week before. Rules:
- SQLI-QS-001 plan regression: new plan >= 2x slower than the best plan
from the week before and >= 60 s extra per day; scripts the force
- SQLI-QS-002 forced plan that fails (e.g. NO_INDEX); scripts the unforce
- SQLI-QS-003 parameter sensitivity: very different plans or volatile
durations; suggests PSP optimization / Query Store hints on 2022+
Also: backups taken in the same second a database was created count.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E3EExwndy8Z22EK2Qpt6BK
- ASP.NET Core Blazor static SSR: overview with KPI tiles, instance page
(findings, CPU/batch/PLE trends, top waits, file latency, blocking,
deadlocks with .xdl download, not-measurable rules, resolved), finding
detail with explanation, evidence, fix script and references
- Light-first design (sidebar, cards, self-hosted Inter, teal accent),
optional dark theme via toggle; strict CSP, no inline styles
- Windows authentication (Negotiate/IIS) with optional group restriction;
refuses to run without authentication outside Development
- Hosting: IIS (in-process ASP.NET Core Module), Windows service with
Kestrel, or systemd; web reads the repository with db_datareader only
- docs/installation.md for repository, permissions, agent and web
- Fix: CPU used by other processes is not derived on Linux (SystemIdle is
unreliable there); deadlocks are not re-reported or stored twice after
an agent restart; Dapper sends datetime2 so time comparisons are exact
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E3EExwndy8Z22EK2Qpt6BK
- Versioned, embedded migrations serialized with sp_getapplock so the agent
and web UI can both migrate at startup
- Findings tracked per rule and subject: first seen, last seen, resolved;
a rule that could not run never resolves its findings
- Time series (waits, file I/O, CPU, counters), blocking chains and
deadlock graphs; idempotent inserts (timestamps truncated to match
datetime2(0)), hourly aggregation and batched retention deletes
- Read API for the web UI; integration tests against SQL Server
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E3EExwndy8Z22EK2Qpt6BK
Collectors (SQL Server 2000+, version-specific SQL, least privilege):
server, configuration, trace flags, databases (settings, backups in UTC,
CHECKDB, VLFs, Query Store, files, tempdb), suspect pages, plan cache,
wait stats and file I/O as deltas with restart/clear detection,
performance counters, CPU from the scheduler monitor ring buffer,
blocking (sleeping head blockers with open transactions), and deadlocks
from system_health continuing from the last file offset.
Rules: configuration (memory, MAXDOP per Microsoft NUMA guidance, cost
threshold, ad hoc bloat, risky options, pending changes), databases
(state, settings, backups, log backups, CHECKDB, suspect pages, VLFs,
autogrowth, max size, Query Store, tempdb), waits (THREADPOOL, signal
waits, locks, I/O, WRITELOG, parallelism, PAGELATCH), I/O latency, CPU,
memory grants, blocking and deadlocks. Missing data makes a rule not
measurable with the collector's reason.
Integration tests run every collector as the least-privilege login and
reproduce real blocking and a real deadlock; green on 2017 and 2022.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E3EExwndy8Z22EK2Qpt6BK
