updates

When converting `rawRows` to in-memory parts, it is possible to do it
without a mutex.

Signed-off-by: Artem Fetishev <rtm@victoriametrics.com>

docs/victoriametrics/Articles.md

@@ -154,6 +154,13 @@ See [our blog](https://victoriametrics.com/blog) for the latest articles written
 * [Why irate from Prometheus doesn't capture spikes](https://valyala.medium.com/why-irate-from-prometheus-doesnt-capture-spikes-45f9896d7832)
 * [VictoriaMetrics: PromQL compliance](https://medium.com/@romanhavronenko/victoriametrics-promql-compliance-d4318203f51e)
 * [How do open source solutions for logs work: Elasticsearch, Loki and VictoriaLogs](https://itnext.io/how-do-open-source-solutions-for-logs-work-elasticsearch-loki-and-victorialogs-9f7097ecbc2f)
+* [How vmagent Collects and Ships Metrics Fast with Aggregation, Deduplication, and More](https://victoriametrics.com/blog/vmagent-how-it-works/)
+* [When Metrics Meet vminsert: A Data-Delivery Story](https://victoriametrics.com/blog/vminsert-how-it-works/)
+* [How vmstorage Handles Data Ingestion From vminsert](https://victoriametrics.com/blog/vmstorage-how-it-handles-data-ingestion/)
+* [How vmstorage Processes Data: Retention, Merging, Deduplication,...](https://victoriametrics.com/blog/vmstorage-retention-merging-deduplication/)
+* [How vmstorage's IndexDB Works](https://victoriametrics.com/blog/vmstorage-how-indexdb-works/)
+* [How vmstorage Handles Query Requests From vmselect](https://victoriametrics.com/blog/vmstorage-how-it-handles-query-requests/)
+* [Inside vmselect: The Query Processing Engine of VictoriaMetrics](https://victoriametrics.com/blog/vmselect-how-it-works/)
 
 ### Tutorials, guides and how-to articles
 
@@ -171,6 +178,12 @@ See [our guides](https://docs.victoriametrics.com/guides/) for the up-to-date gu
 * [Prometheus storage: tech terms for humans](https://valyala.medium.com/prometheus-storage-technical-terms-for-humans-4ab4de6c3d48)
 * [Cardinality explorer](https://victoriametrics.com/blog/cardinality-explorer/)
 * [Rules backfilling via vmalert](https://victoriametrics.com/blog/rules-replay/)
+* [vmagent: Key Features Explained in Under 15 Minutes](https://victoriametrics.com/blog/vmagent-key-features-explained/)
+* [Prometheus Metrics Explained: Counters, Gauges, Histograms & Summaries](https://victoriametrics.com/blog/prometheus-monitoring-metrics-counters-gauges-histogram-summaries/)
+* [Prometheus Monitoring: Instant Queries and Range Queries Explained](https://victoriametrics.com/blog/prometheus-monitoring-instant-range-query/)
+* [Prometheus Monitoring: Functions, Subqueries, Operators, and Modifiers](https://victoriametrics.com/blog/prometheus-monitoring-function-operator-modifier/)
+* [Prometheus Alerting 101: Rules, Recording Rules, and Alertmanager](https://victoriametrics.com/blog/alerting-recording-rules-alertmanager/)
+* [Alerting Best Practices](https://victoriametrics.com/blog/alerting-best-practices/)
 
 ### Other articles
 

docs/victoriametrics/Cluster-VictoriaMetrics.md

@@ -57,6 +57,10 @@ Each service may scale independently and may run on the most suitable hardware.
 This is a [shared nothing architecture](https://en.wikipedia.org/wiki/Shared-nothing_architecture).
 It increases cluster availability, and simplifies cluster maintenance as well as cluster scaling.
 
+> Further reading, deep dives into how each service works internally:
+> - `vmstorage`: [How vmstorage Handles Data Ingestion From vminsert](https://victoriametrics.com/blog/vmstorage-how-it-handles-data-ingestion/), [How vmstorage's IndexDB Works](https://victoriametrics.com/blog/vmstorage-how-indexdb-works/), [How vmstorage Handles Query Requests From vmselect](https://victoriametrics.com/blog/vmstorage-how-it-handles-query-requests/).
+> - `vmselect`: [Inside vmselect: The Query Processing Engine of VictoriaMetrics](https://victoriametrics.com/blog/vmselect-how-it-works/).
+
 ![Cluster Scheme](Cluster-VictoriaMetrics-components.webp)
 
 ## vmui
@@ -834,7 +838,7 @@ This ensures that incoming metrics are evenly distributed across all `vmstorage`
 The downside is that a single slow vmstorage node can throttle the entire cluster.
 
 When `-disableRerouting=false` is enabled on `vminsert`, 
-the cluster will automatically re-route writes away from the slowest vmstorage node to preserve maximum ingestion throughput.
+the cluster will automatically [re-route writes](https://victoriametrics.com/blog/vminsert-how-it-works/#31-rerouting) away from the slowest vmstorage node to preserve maximum ingestion throughput.
 
 Re-routing occurs only when all of the following conditions hold:
 - the storage send buffer is full.
@@ -877,7 +881,7 @@ See also [resource usage limits docs](#resource-usage-limits).
 
 ## Rebalancing
 
-Every `vminsert` node evenly spreads (shards) incoming data among `vmstorage` nodes specified in the `-storageNode` command-line flag.
+Every `vminsert` node [evenly spreads (shards) incoming data](https://victoriametrics.com/blog/vminsert-how-it-works/#3-sharding-and-buffering) among `vmstorage` nodes specified in the `-storageNode` command-line flag.
 This guarantees even distribution of the ingested data among `vmstorage` nodes. When new `vmstorage` nodes are added to the `-storageNode`
 command-line flag at `vminsert`, then only newly ingested data is distributed evenly among old and new `vmstorage` nodes, while
 historical data remains on the old `vmstorage` nodes. This speeds up data ingestion and querying for the majority of production workloads,
@@ -1025,7 +1029,7 @@ By default, VictoriaMetrics offloads replication to the underlying storage point
 which guarantees data durability. VictoriaMetrics supports application-level replication if replicated durable persistent disks cannot be used for some reason.
 
 The replication can be enabled by passing `-replicationFactor=N` command-line flag to `vminsert`. This instructs `vminsert` to store `N` copies for every ingested sample
-on `N` distinct `vmstorage` nodes. This guarantees that all the stored data remains available for querying if up to `N-1` `vmstorage` nodes are unavailable.
+on `N` distinct `vmstorage` nodes. This guarantees that all the stored data remains available for querying if up to `N-1` `vmstorage` nodes are unavailable. See [how `vminsert` replicates each sample to `N` `vmstorage` nodes](https://victoriametrics.com/blog/vminsert-how-it-works/#4-replication-and-sending-data-to-vmstorage) for details.
 
 Passing `-replicationFactor=N` command-line flag to `vmselect` instructs it to not mark responses as `partial` if less than `-replicationFactor` vmstorage nodes are unavailable during the query.
 See [cluster availability docs](#cluster-availability) for details.
@@ -1060,7 +1064,7 @@ deduplication can't be guaranteed when samples and sample duplicates for the sam
 - when `vmstorage` node has no enough capacity for processing incoming data stream. Then `vminsert` re-routes new samples to other `vmstorage` nodes.
 
 It is recommended to set **the same** `-dedup.minScrapeInterval` command-line flag value to both `vmselect` and `vmstorage` nodes
-to ensure query results consistency, even if storage layer didn't complete deduplication yet.
+to ensure query results consistency, even if [storage layer didn't complete deduplication](https://victoriametrics.com/blog/vmstorage-retention-merging-deduplication/#deduplication) yet.
 
 ## Metrics Metadata
 

docs/victoriametrics/MetricsQL.md

@@ -59,6 +59,7 @@ Other PromQL functionality should work the same in MetricsQL.
 
 MetricsQL implements [PromQL](https://medium.com/@valyala/promql-tutorial-for-beginners-9ab455142085)
 and provides additional functionality mentioned below, which is aimed towards solving practical cases.
+See [operators and modifiers in MetricsQL](https://victoriametrics.com/blog/prometheus-monitoring-function-operator-modifier/#operators--modifiers) for details.
 Feel free [filing a feature request](https://github.com/VictoriaMetrics/VictoriaMetrics/issues) if you think MetricsQL misses certain useful functionality.
 
 This functionality can be evaluated at [VictoriaMetrics demo playground](https://play.victoriametrics.com/select/accounting/1/6a716b0f-38bc-4856-90ce-448fd713e3fe/prometheus/graph/)
@@ -156,7 +157,7 @@ MetricsQL provides the following functions:
 
 ### Rollup functions
 
-**Rollup functions** (aka range functions or window functions) calculate rollups over [raw samples](https://docs.victoriametrics.com/victoriametrics/keyconcepts/#raw-samples)
+**[Rollup functions](https://victoriametrics.com/blog/prometheus-monitoring-function-operator-modifier/#rollup-functions)** (aka range functions or window functions) calculate rollups over [raw samples](https://docs.victoriametrics.com/victoriametrics/keyconcepts/#raw-samples)
 on the given lookbehind window for the [selected time series](https://docs.victoriametrics.com/victoriametrics/keyconcepts/#filtering).
 For example, `avg_over_time(temperature[24h])` calculates the average temperature over [raw samples](https://docs.victoriametrics.com/victoriametrics/keyconcepts/#raw-samples) for the last 24 hours.
 
@@ -348,6 +349,7 @@ If the lookbehind window is skipped in square brackets, then it is automatically
 passed to [/api/v1/query_range](https://docs.victoriametrics.com/victoriametrics/keyconcepts/#range-query) or [/api/v1/query](https://docs.victoriametrics.com/victoriametrics/keyconcepts/#instant-query),
 while `scrape_interval` is the interval between [raw samples](https://docs.victoriametrics.com/victoriametrics/keyconcepts/#raw-samples) for the selected time series.
 This allows avoiding unexpected gaps on the graph when `step` is smaller than the `scrape_interval`.
+See [range vector selectors](https://victoriametrics.com/blog/prometheus-monitoring-instant-range-query/#range-vector-selector) for details.
 
 #### delta
 
@@ -1110,7 +1112,7 @@ See also [zscore](#zscore), [range_trim_zscore](#range_trim_zscore) and [outlier
 
 ### Transform functions
 
-**Transform functions** calculate transformations over [rollup results](#rollup-functions).
+**[Transform functions](https://victoriametrics.com/blog/prometheus-monitoring-function-operator-modifier/#transformation-functions)** calculate transformations over [rollup results](#rollup-functions).
 For example, `abs(delta(temperature[24h]))` calculates the absolute value for every point of every time series
 returned from the rollup `delta(temperature[24h])`.
 
@@ -1846,7 +1848,7 @@ This function is supported by PromQL.
 
 ### Label manipulation functions
 
-**Label manipulation functions** perform manipulations with labels on the selected [rollup results](#rollup-functions).
+**[Label manipulation functions](https://victoriametrics.com/blog/prometheus-monitoring-function-operator-modifier/#label-manipulation-functions)** perform manipulations with labels on the selected [rollup results](#rollup-functions).
 
 Additional details:
 
@@ -2016,7 +2018,7 @@ See also [sort_by_label_numeric](#sort_by_label_numeric) and [sort_by_label_desc
 
 ### Aggregate functions
 
-**Aggregate functions** calculate aggregates over groups of [rollup results](#rollup-functions).
+**[Aggregate functions](https://victoriametrics.com/blog/prometheus-monitoring-function-operator-modifier/#aggregation-functions)** calculate aggregates over groups of [rollup results](#rollup-functions).
 
 Additional details:
 
@@ -2339,7 +2341,7 @@ See also [zscore_over_time](#zscore_over_time), [range_trim_zscore](#range_trim_
 
 ## Subqueries
 
-MetricsQL supports and extends PromQL subqueries. See [this article](https://valyala.medium.com/prometheus-subqueries-in-victoriametrics-9b1492b720b3) for details.
+MetricsQL supports and extends PromQL [subqueries](https://victoriametrics.com/blog/prometheus-monitoring-function-operator-modifier/#subqueries). See [this article](https://valyala.medium.com/prometheus-subqueries-in-victoriametrics-9b1492b720b3) for details.
 Any [rollup function](#rollup-functions) for something other than [series selector](https://docs.victoriametrics.com/victoriametrics/keyconcepts/#filtering) form a subquery.
 Nested rollup functions can be implicit thanks to the [implicit query conversions](#implicit-query-conversions).
 For example, `delta(sum(m))` is implicitly converted to `delta(sum(default_rollup(m))[1i:1i])`, so it becomes a subquery,

docs/victoriametrics/README.md

@@ -1314,6 +1314,7 @@ per each `-dedup.minScrapeInterval` discrete interval if `-dedup.minScrapeInterv
 For example, `-dedup.minScrapeInterval=60s` would leave a single raw sample with the biggest timestamp per each discrete
 `60s` interval.
 This aligns with the [staleness rules in Prometheus](https://prometheus.io/docs/prometheus/latest/querying/basics/#staleness).
+See [how deduplication works](https://victoriametrics.com/blog/vmstorage-retention-merging-deduplication/#deduplication) for details.
 
 If multiple raw samples have **the same timestamp** on the given `-dedup.minScrapeInterval` discrete interval,
 then the sample with **the biggest value** is kept.
@@ -1397,7 +1398,7 @@ in separate files under `part` directory - `timestamps.bin` and `values.bin`.
 The `part` directory also contains `index.bin` and `metaindex.bin` files - these files contain index
 for fast block lookups, which belong to the given `TSID` and cover the given time range.
 
-`Parts` are periodically merged into bigger parts in background. The background merge provides the following benefits:
+`Parts` are periodically merged into bigger parts in background. The [background merge](https://victoriametrics.com/blog/vmstorage-retention-merging-deduplication/#merge-process) provides the following benefits:
 
 * keeping the number of data files under control, so they don't exceed limits on open files
 * improved data compression, since bigger parts are usually compressed better than smaller parts
@@ -1532,6 +1533,7 @@ are **eventually deleted** during [background merge](https://medium.com/@valyala
 The time range covered by data part is **not limited by retention period unit**. One data part can cover hours or days of
 data. Hence, a data part can be deleted only **when fully outside the configured retention**.
 See more about partitions and parts in the [Storage section](#storage).
+See [how retention frees disk space](https://victoriametrics.com/blog/vmstorage-retention-merging-deduplication/#retention-free-disk-space-guard-and-downsampling) for details.
 
 The maximum disk space usage for a given `-retentionPeriod` is going to be (`-retentionPeriod` + 1) months.
 For example, if `-retentionPeriod` is set to 1, data for January is deleted on March 1st.
@@ -1634,6 +1636,7 @@ See how to request a [free trial license](https://victoriametrics.com/products/e
 This command-line flag instructs leaving the last sample per each `interval` for [time series](https://docs.victoriametrics.com/victoriametrics/keyconcepts/#time-series)
 [samples](https://docs.victoriametrics.com/victoriametrics/keyconcepts/#raw-samples) older than the `offset`. The `offset` must be a multiple of `interval`. For example, `-downsampling.period=30d:5m` instructs leaving the last sample
 per each 5-minute interval for samples older than 30 days, while the rest of samples are dropped.
+See [Enterprise downsampling internals](https://victoriametrics.com/blog/vmstorage-retention-merging-deduplication/#retention-filters-and-downsampling-enterprise-plan) for details.
 
 The `-downsampling.period` command-line flag can be specified multiple times in order to apply different downsampling levels for different time ranges (aka multi-level downsampling).
 For example, `-downsampling.period=30d:5m,180d:1h` instructs leaving the last sample per each 5-minute interval for samples older than 30 days,

docs/victoriametrics/changelog/CHANGELOG.md

@@ -25,7 +25,6 @@ The sandbox cluster installation runs under the constant load generated by
 See also [LTS releases](https://docs.victoriametrics.com/victoriametrics/lts-releases/).
 
 ## tip
-* BUGFIX: [vmagent](https://docs.victoriametrics.com/victoriametrics/vmagent/): flush pending persistent queue data to chunk file before updating the metadata. This prevents the metadata writer offset from getting ahead of the chunk file size and avoids losing the persistent queue after an unclean shutdown. See [#11192](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/11192).
 
 * FEATURE: [MetricsQL](https://docs.victoriametrics.com/victoriametrics/metricsql/): support `fill` modifiers to allow missing series on either side of a binary operation to be filled with a provided default value. See [#10598](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/10598).
 * FEATURE: [vmagent](https://docs.victoriametrics.com/victoriametrics/vmagent/) and [vmsingle](https://docs.victoriametrics.com/victoriametrics/single-server-victoriametrics/): Improve background discovery performance for [http_sd](https://docs.victoriametrics.com/victoriametrics/sd_configs/#http_sd_configs) discovery. See [#8838](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/8838).
@@ -37,6 +36,7 @@ See also [LTS releases](https://docs.victoriametrics.com/victoriametrics/lts-rel
 * BUGFIX: `vmselect` in [VictoriaMetrics cluster](https://docs.victoriametrics.com/victoriametrics/cluster-victoriametrics/): properly apply limit to metrics metadata response. See [#11139](https://github.com/VictoriaMetrics/VictoriaMetrics/pull/11139).
 * BUGFIX: [vmui](https://docs.victoriametrics.com/victoriametrics/single-server-victoriametrics/#vmui): keep only one header navigation dropdown (`Explore`, `Tools`) open at a time. Previously, hovering across two dropdowns could briefly leave both open due to the close delay. See [#11224](https://github.com/VictoriaMetrics/VictoriaMetrics/pull/11224). Thanks to @antedotee for contribution.
 * BUGFIX: [vmagent](https://docs.victoriametrics.com/victoriametrics/vmagent/): fix a possible data race when processing OpenTelemetry metadata. See [#11238](https://github.com/VictoriaMetrics/VictoriaMetrics/pull/11238). Thanks to @nevgeny for contribution.
+* BUGFIX: [vmagent](https://docs.victoriametrics.com/victoriametrics/vmagent/): flush pending persistent queue data to chunk file before updating the metadata. This prevents the metadata writer offset from getting ahead of the chunk file size and avoids losing the persistent queue after an unclean shutdown. See [#11192](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/11192).
 
 ## [v1.147.0](https://github.com/VictoriaMetrics/VictoriaMetrics/releases/tag/v1.147.0)
 

docs/victoriametrics/keyConcepts/keyConcepts.md

@@ -40,7 +40,7 @@ requests_total{path="/", code="403"}
 The meta-information - a set of `labels` in curly braces - gives us a context for which `path` and with what `code`
 the `request` was served. Label-value pairs are always of a `string` type. VictoriaMetrics data model is schemaless,
 which means there is no need to define metric names or their labels in advance. The user is free to add or change ingested
-metrics anytime.
+metrics anytime. See [metric names and labels explained](https://victoriametrics.com/blog/prometheus-monitoring-metrics-counters-gauges-histogram-summaries/#metric-name-and-labels) for details.
 
 Actually, the metric name is also a label with a special name `__name__`. 
 The `__name__` key could be omitted {{% available_from "v1.111.0" %}} for simplicity. So the following series are identical:
@@ -75,7 +75,7 @@ See [these docs](https://docs.victoriametrics.com/victoriametrics/faq/#what-is-h
 
 #### Raw samples
 
-Every unique time series may consist of an arbitrary number of `(value, timestamp)` data points (aka `raw samples`) sorted by `timestamp`.
+Every unique time series may consist of an arbitrary number of `(value, timestamp)` data points (aka [`raw samples`](https://victoriametrics.com/blog/prometheus-monitoring-metrics-counters-gauges-histogram-summaries/#sample)) sorted by `timestamp`.
 VictoriaMetrics stores all the `values` as [float64](https://en.wikipedia.org/wiki/Double-precision_floating-point_format)
 with [extra compression](https://faun.pub/victoriametrics-achieving-better-compression-for-time-series-data-than-gorilla-317bc1f95932) applied.
 This allows storing precise integer values with up to 12 decimal digits and any floating-point values with up to 12 significant decimal digits.
@@ -126,7 +126,7 @@ type exists specifically to help users to understand how the metric was measured
 
 #### Counter
 
-Counter is a metric, which counts some events. Its value increases or stays the same over time.
+[Counter](https://victoriametrics.com/blog/prometheus-monitoring-metrics-counters-gauges-histogram-summaries/#counter) is a metric, which counts some events. Its value increases or stays the same over time.
 It cannot decrease in general case. The only exception is e.g. `counter reset`,
 when the metric resets to zero. The `counter reset` can occur when the service, which exposes the counter, restarts.
 So, the `counter` metric shows the number of observed events since the service start.
@@ -157,7 +157,7 @@ by humans from other metric types.
 
 #### Gauge
 
-Gauge is used for measuring a value that can go up and down:
+[Gauge](https://victoriametrics.com/blog/prometheus-monitoring-metrics-counters-gauges-histogram-summaries/#gauge) is used for measuring a value that can go up and down:
 
 ![gauge](gauge.webp)
 
@@ -178,7 +178,7 @@ and [rollup functions](https://docs.victoriametrics.com/victoriametrics/metricsq
 
 #### Histogram
 
-Histogram is a set of [counter](#counter) metrics with different `vmrange` or `le` labels.
+[Histogram](https://victoriametrics.com/blog/prometheus-monitoring-metrics-counters-gauges-histogram-summaries/#histogram) is a set of [counter](#counter) metrics with different `vmrange` or `le` labels.
 The `vmrange` or `le` labels define measurement boundaries of a particular bucket.
 When the observed measurement hits a particular bucket, then the corresponding counter is incremented.
 
@@ -282,7 +282,7 @@ We recommend reading the following articles before you start using histograms:
 
 #### Summary
 
-Summary metric type is quite similar to [histogram](#histogram) and is used for
+[Summary](https://victoriametrics.com/blog/prometheus-monitoring-metrics-counters-gauges-histogram-summaries/#summary) metric type is quite similar to [histogram](#histogram) and is used for
 [quantiles](https://prometheus.io/docs/practices/histograms/#quantiles) calculations. The main difference
 is that calculations are made on the client-side, so metrics exposition format already contains pre-defined
 quantiles:
@@ -504,6 +504,7 @@ Params:
 The result of Instant query is a list of [time series](https://docs.victoriametrics.com/victoriametrics/keyconcepts/#time-series)
 matching the filter in `query` expression. Each returned series contains exactly one `(timestamp, value)` entry,
 where `timestamp` equals to the `time` query arg, while the `value` contains `query` result at the requested `time`.
+See [instant vectors explained](https://victoriametrics.com/blog/prometheus-monitoring-instant-range-query/#instant-vector) for details.
 
 To understand how instant queries work, let's begin with a data sample:
 
@@ -606,6 +607,7 @@ at `start`, `start+step`, `start+2*step`, ..., `start+N*step` timestamps. In oth
 executed independently at `start`, `start+step`, ..., `start+N*step` timestamps with the only difference that an instant query
 does not return `ephemeral` samples (see below). Instead, if the database does not contain any samples for the requested time and step,
 it simply returns an empty result.
+See [range vectors explained](https://victoriametrics.com/blog/prometheus-monitoring-instant-range-query/#range-vector) for details.
 
 
 For example, to get the values of `foo_bar` during the time range from `2022-05-10T07:59:00Z` to `2022-05-10T08:17:00Z`,

docs/victoriametrics/vmagent.md

@@ -127,7 +127,8 @@ See [these docs](#how-to-collect-metrics-in-prometheus-format) for details.
 
 `vmagent` can be used as an alternative to [StatsD](https://github.com/statsd/statsd)
 when [stream aggregation](https://docs.victoriametrics.com/victoriametrics/stream-aggregation/) is enabled.
-See [these docs](https://docs.victoriametrics.com/victoriametrics/stream-aggregation/#statsd-alternative) for details.
+See [these docs](https://docs.victoriametrics.com/victoriametrics/stream-aggregation/#statsd-alternative) for details,
+or the blog post on [how vmagent's stream aggregation works](https://victoriametrics.com/blog/vmagent-key-features-explained/#stream-aggregation).
 
 ### Flexible metrics relay
 
@@ -137,7 +138,7 @@ to other remote storage systems that support Prometheus `remote_write` protocol
 
 ### Replication and high availability
 
-`vmagent` replicates the collected metrics among multiple remote storage instances configured via `-remoteWrite.url` args.
+`vmagent` [replicates the collected metrics](https://victoriametrics.com/blog/vmagent-how-it-works/#step-4-sharding--replication) among multiple remote storage instances configured via `-remoteWrite.url` args.
 If a single remote storage instance is temporarily unavailable, the collected data remains available on the other remote storage instances.
 `vmagent` buffers the collected data in files at `-remoteWrite.tmpDataPath` until the remote storage becomes available again.
 Then it sends the buffered data to the remote storage in order to prevent data gaps.
@@ -150,7 +151,8 @@ See [these docs](https://docs.victoriametrics.com/victoriametrics/cluster-victor
 
 `vmagent` can add, remove, or update labels on the collected data before sending it to the remote storage.
 It can filter scrape targets or remove unwanted samples via Prometheus-like relabeling.
-Please see the [Relabeling cookbook](https://docs.victoriametrics.com/victoriametrics/relabeling/) for details.
+Please see the [Relabeling cookbook](https://docs.victoriametrics.com/victoriametrics/relabeling/) for details,
+or the blog post on [how vmagent applies global relabeling to reduce cardinality](https://victoriametrics.com/blog/vmagent-how-it-works/#step-2-global-relabeling-cardinality-reduction).
 
 ### Sharding among remote storages
 
@@ -158,6 +160,7 @@ By default, `vmagent` replicates data to remote storage systems via the `-remote
 If the `-remoteWrite.shardByURL` command-line flag is set, then `vmagent` spreads
 the outgoing [time series](https://docs.victoriametrics.com/victoriametrics/keyconcepts/#time-series) evenly among all the remote storage
 systems listed in `-remoteWrite.url`.
+See the blog post on [how vmagent shards data across remote storage systems](https://victoriametrics.com/blog/vmagent-key-features-explained/#sharding) for the details.
 
 It is possible to replicate samples among remote storage systems by passing `-remoteWrite.shardByURLReplicas=N`
 to `vmagent` in addition to the `-remoteWrite.shardByURL` command-line flag.
@@ -268,7 +271,9 @@ for the collected samples. Examples:
   ```sh
   ./vmagent -remoteWrite.url=http://remote-storage/api/v1/write -streamAggr.dropInputLabels=replica -streamAggr.dedupInterval=60s
   ```
-  
+
+See the blog post on [how vmagent performs global deduplication and stream aggregation](https://victoriametrics.com/blog/vmagent-how-it-works/#step-3-global-deduplication--stream-aggregation) for the details.
+
 ### Monitoring Data eXchange
 
 The Monitoring Data eXchange (MDX){{% available_from "v1.147.0" %}} feature allows `vmagent` to forward only VictoriaMetrics metrics to selected `-remoteWrite.url` destinations while dropping metrics from non-VictoriaMetrics services.
@@ -357,6 +362,8 @@ in addition to the pull-based Prometheus-compatible targets' scraping:
 * Prometheus exposition format via `http://<vmagent>:8429/api/v1/import/prometheus`. See [these docs](https://docs.victoriametrics.com/victoriametrics/single-server-victoriametrics/#how-to-import-data-in-prometheus-exposition-format) for details.
 * Arbitrary CSV data via `http://<vmagent>:8429/api/v1/import/csv`. See [these docs](https://docs.victoriametrics.com/victoriametrics/single-server-victoriametrics/#how-to-import-csv-data).
 
+See the blog post on [how vmagent receives data via push APIs or scraping](https://victoriametrics.com/blog/vmagent-how-it-works/#step-1-receiving-data-via-api-or-scrape) for the details.
+
 ## How to collect metrics in Prometheus format
 
 Specify the path to the `prometheus.yml` file via the `-promscrape.config` command-line flag. `vmagent` takes into account the following
@@ -490,15 +497,13 @@ SRV URLs are useful when HTTP services run on different TCP ports or when their
 
 When comparing the remote protocols between VictoriaMetrics and Prometheus, VictoriaMetrics provides the following benefits:
 
-* Reduced network bandwidth usage by 2x-5x. This allows saving network bandwidth usage costs when `vmagent` and
+* Reduced network bandwidth usage by 2x-5x. This allows [saving network bandwidth usage costs](https://victoriametrics.com/blog/victoriametrics-remote-write/) when `vmagent` and
   the configured remote storage systems are located in different datacenters, availability zones, or regions.
 
 * Reduced disk read/write IO and disk space usage at `vmagent` when the remote storage is temporarily unavailable.
   In this case, `vmagent` buffers incoming data to disk using the VictoriaMetrics remote write format.
   This reduces disk read/write IO and disk space usage by 2x-5x compared to the Prometheus remote write format.
 
-> See blogpost [Save network costs with VictoriaMetrics remote write protocol](https://victoriametrics.com/blog/victoriametrics-remote-write/).
-
 `vmagent` uses VictoriaMetrics remote write protocol by default {{% available_from "v1.116.0" %}} when it sends data to VictoriaMetrics components such as other `vmagent` instances,
 [single-node VictoriaMetrics](https://docs.victoriametrics.com/victoriametrics/single-server-victoriametrics/)
 or `vminsert` at [cluster version](https://docs.victoriametrics.com/victoriametrics/cluster-victoriametrics/). If needed, it can automatically downgrade to a Prometheus protocol at runtime.
@@ -835,7 +840,7 @@ as soon as it is parsed in stream parsing mode.
 ## Scraping big number of targets
 
 A single `vmagent` instance can scrape tens of thousands of scrape targets. Sometimes this isn't enough due to limitations on CPU, network, RAM, etc.
-In this case, scrape targets can be split among multiple `vmagent` instances (aka `vmagent` horizontal scaling, sharding, and clustering).
+In this case, scrape targets can be split among multiple `vmagent` instances (aka [`vmagent` horizontal scaling](https://victoriametrics.com/blog/vmagent-key-features-explained/#scaling-vmagent), sharding, and clustering).
 The number of `vmagent` instances in the cluster must be passed to the `-promscrape.cluster.membersCount` command-line flag.
 Each `vmagent` instance in the cluster must use identical `-promscrape.config` files with distinct `-promscrape.cluster.memberNum` values
 in the range `0 ... N-1`, where `N` is the number of `vmagent` instances in the cluster specified via `-promscrape.cluster.membersCount`.
@@ -948,7 +953,7 @@ scrape_configs:
 
 ## On-disk persistence
 
-`vmagent` stores pending data that cannot be sent to the configured remote storage systems in a timely manner.
+`vmagent` [stores pending data that cannot be sent to the configured remote storage systems](https://victoriametrics.com/blog/vmagent-key-features-explained/#persistent-disk-for-remote-write) in a timely manner.
 By default, `vmagent` writes all the pending data to the folder configured via `-remoteWrite.tmpDataPath` cmd-line flag
 until this data is sent to the configured `-remoteWrite.url` systems or until the folder becomes full.
 The maximum data size that can be saved to `-remoteWrite.tmpDataPath` per every configured `-remoteWrite.url` can be
@@ -996,7 +1001,7 @@ moment it becomes visible at the remote storage.
 
 This behavior can be changed with the `-remoteWrite.inmemoryQueues` {{% available_from "v1.146.0" %}} command-line flag.
 When set to a non-zero value, vmagent starts the given number of additional workers,
-which send only recently ingested data from the in-memory queue, while the workers configured via `-remoteWrite.queues` drain the file-based backlog concurrently.
+which send only recently ingested data from the [in-memory queue](https://victoriametrics.com/blog/vmagent-how-it-works/#in-memory-queue), while the workers configured via `-remoteWrite.queues` drain the file-based backlog concurrently.
 This reduces the delivery lag for fresh samples after remote storage outages or slowdowns. The flag can be set individually per each `-remoteWrite.url`.
 
 Note that these workers are started in addition to the workers configured via `-remoteWrite.queues`, so the total number of concurrent connections to

docs/victoriametrics/vmalert.md

@@ -209,6 +209,8 @@ rules:
 [ debug: <bool> | default = false ]
 ```
 
+See a blogpost about [how rule groups work](https://victoriametrics.com/blog/alerting-recording-rules-alertmanager/#groups).
+
 ### Rules
 
 Every rule contains an `expr` field for the expression to evaluate against the configured datasource.
@@ -285,6 +287,8 @@ annotations:
   [ <labelname>: <tmpl_string> ]
 ```
 
+See a blogpost about [how alerting rules work](https://victoriametrics.com/blog/alerting-recording-rules-alertmanager/#alerting-rules).
+
 #### Recording rules
 
 The syntax for recording rules is the following:
@@ -322,10 +326,13 @@ labels:
 
 For recording rules to work `-remoteWrite.url` must be specified.
 
+See a blogpost about [how recording rules work](https://victoriametrics.com/blog/alerting-recording-rules-alertmanager/#recording-rules).
+
 ## Templating
 
 It is allowed to use [Go templating](https://golang.org/pkg/text/template/) in annotations and labels(with limited support) to format data, iterate over
 or execute expressions.
+See a blogpost about [templating alerts](https://victoriametrics.com/blog/alerting-recording-rules-alertmanager/#templates).
 The following variables are available in templating:
 
 | Variable                           | Description                                                                                                                                           | Example                                                                                                                                                                                |
@@ -928,7 +935,7 @@ See full description for these flags in `./vmalert -help`.
 
 ## Unit Testing for Rules
 
-You can use `vmalert-tool` to test your alerting and recording rules like [promtool does](https://prometheus.io/docs/prometheus/latest/configuration/unit_testing_rules/).
+You can use `vmalert-tool` to [test your alerting and recording rules](https://victoriametrics.com/blog/alerting-best-practices/#testing-alerts) like [promtool does](https://prometheus.io/docs/prometheus/latest/configuration/unit_testing_rules/).
 See more details in [vmalert-tool](https://docs.victoriametrics.com/victoriametrics/vmalert-tool/#unit-testing-for-rules).
 
 ## Monitoring
@@ -982,6 +989,8 @@ Try the following tips to avoid common issues:
     In that case, the default step will be used (`-datasource.queryStep`) and may cause unexpected results compared to
     executing this query in vmui/Grafana, where step is adjusted differently.
 
+See a blogpost about [reducing alert noise](https://victoriametrics.com/blog/alerting-best-practices/#reducing-noise).
+
 ### Rule state
 
 vmalert keeps the last `-rule.updateEntriesLimit` updates (or `update_entries_limit` [per-rule config](https://docs.victoriametrics.com/victoriametrics/vmalert/#alerting-rules))
@@ -1037,6 +1046,8 @@ Sometimes, it's hard to understand why a specific alert fired or not. Keep in mi
 If evaluation returns error (i.e. datasource is unavailable), alert state doesn't change.
 If at least one evaluation returns no data, then alert's `for` state resets.
 
+See a blogpost about [tuning the `for` param](https://victoriametrics.com/blog/alerting-best-practices/#the-for-param).
+
 > Note: The alert state is tracked separately for each time series returned during evaluation.
 > For example, if the 1st evaluation returns series A and B, and the 2nd evaluation returns only B – the alert will remain active **only for B**.
 
@@ -1110,6 +1121,8 @@ How to reduce the chance for a rule to flap:
 
 See [common mistakes](#common-mistakes) for rules config.
 
+See a blogpost about [tuning `keep_firing_for`](https://victoriametrics.com/blog/alerting-best-practices/#the-keep_firing_for-param).
+
 ### Never-firing alerts
 
 vmalert can detect {{% available_from "v1.91.0" %}} if alert's expression doesn't match any time series in runtime.
@@ -1360,6 +1373,8 @@ The list of configured or discovered Notifiers can be explored via [UI](#web).
 If Alertmanager runs in cluster mode then all its URLs needs to be available during discovery
 to ensure [high availability](https://github.com/prometheus/alertmanager#high-availability).
 
+See a blogpost about [how Alertmanager processes alerts](https://victoriametrics.com/blog/alerting-recording-rules-alertmanager/#alertmanager).
+
 The configuration file [specification](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmalert/notifier/config.go)
 is the following:
 

lib/storage/partition.go

@@ -11,9 +11,7 @@ import (
 	"sync"
 	"sync/atomic"
 	"time"
-	"unsafe"
 
-	"github.com/VictoriaMetrics/VictoriaMetrics/lib/atomicutil"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/cgroup"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/encoding"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/fs"
@@ -40,14 +38,6 @@ const maxInmemoryParts = 60
 // See appendPartsToMerge tests for details.
 const defaultPartsToMerge = 15
 
-// The number of shards for rawRow entries per partition.
-//
-// Higher number of shards reduces CPU contention and increases the max bandwidth on multi-core systems.
-var rawRowsShardsPerPartition = cgroup.AvailableCPUs()
-
-// The interval for flushing buffered rows into parts, so they become visible to search.
-const pendingRowsFlushInterval = 2 * time.Second
-
 // The interval for guaranteed flush of recently ingested data from memory to on-disk parts, so they survive process crash.
 var dataFlushInterval = 5 * time.Second
 
@@ -66,11 +56,6 @@ func SetDataFlushInterval(d time.Duration) {
 	dataFlushInterval = d
 }
 
-// The maximum number of rawRow items in rawRowsShard.
-//
-// Limit the maximum shard size to 8Mb, since this gives the lowest CPU usage under high ingestion rate.
-const maxRawRowsPerShard = (8 << 20) / int(unsafe.Sizeof(rawRow{}))
-
 // partition represents a partition.
 type partition struct {
 	activeInmemoryMerges atomic.Int64
@@ -480,150 +465,29 @@ func (pt *partition) AddRows(rows []rawRow) {
 		}
 	}
 
-	pt.rawRows.addRows(pt, rows)
+	pt.rawRows.addRows(pt.flushRowssToInmemoryParts, rows)
 }
 
 var isDebug = false
 
-type rawRowsShards struct {
-	flushDeadlineMs atomic.Int64
-
-	shardIdx atomic.Uint32
-
-	// Shards reduce lock contention when adding rows on multi-CPU systems.
-	shards []rawRowsShard
-
-	rowssToFlushLock sync.Mutex
-	rowssToFlush     [][]rawRow
-}
-
-func (rrss *rawRowsShards) init() {
-	rrss.shards = make([]rawRowsShard, rawRowsShardsPerPartition)
-}
-
-func (rrss *rawRowsShards) addRows(pt *partition, rows []rawRow) {
-	shards := rrss.shards
-	shardsLen := uint32(len(shards))
-	for len(rows) > 0 {
-		n := rrss.shardIdx.Add(1)
-		idx := n % shardsLen
-		tailRows, rowsToFlush := shards[idx].addRows(rows)
-		rrss.addRowsToFlush(pt, rowsToFlush)
-		rows = tailRows
-	}
-}
-
-func (rrss *rawRowsShards) addRowsToFlush(pt *partition, rowsToFlush []rawRow) {
-	if len(rowsToFlush) == 0 {
-		return
-	}
-
-	var rowssToMerge [][]rawRow
-
-	rrss.rowssToFlushLock.Lock()
-	if len(rrss.rowssToFlush) == 0 {
-		rrss.updateFlushDeadline()
-	}
-	rrss.rowssToFlush = append(rrss.rowssToFlush, rowsToFlush)
-	if len(rrss.rowssToFlush) >= defaultPartsToMerge {
-		rowssToMerge = rrss.rowssToFlush
-		rrss.rowssToFlush = nil
-	}
-	rrss.rowssToFlushLock.Unlock()
-
-	pt.flushRowssToInmemoryParts(rowssToMerge)
-}
-
-func (rrss *rawRowsShards) Len() int {
-	n := 0
-	for i := range rrss.shards[:] {
-		n += rrss.shards[i].Len()
-	}
-
-	rrss.rowssToFlushLock.Lock()
-	for _, rows := range rrss.rowssToFlush {
-		n += len(rows)
-	}
-	rrss.rowssToFlushLock.Unlock()
-
-	return n
-}
-
-func (rrss *rawRowsShards) updateFlushDeadline() {
-	rrss.flushDeadlineMs.Store(time.Now().Add(pendingRowsFlushInterval).UnixMilli())
-}
-
-type rawRowsShardNopad struct {
-	flushDeadlineMs atomic.Int64
-
-	mu   sync.Mutex
-	rows []rawRow
-}
-
-type rawRowsShard struct {
-	rawRowsShardNopad
-
-	// The padding prevents false sharing
-	_ [atomicutil.CacheLineSize - unsafe.Sizeof(rawRowsShardNopad{})%atomicutil.CacheLineSize]byte
-}
-
-func (rrs *rawRowsShard) Len() int {
-	rrs.mu.Lock()
-	n := len(rrs.rows)
-	rrs.mu.Unlock()
-	return n
-}
-
-func (rrs *rawRowsShard) addRows(rows []rawRow) ([]rawRow, []rawRow) {
-	var rowsToFlush []rawRow
-
-	rrs.mu.Lock()
-	if cap(rrs.rows) == 0 {
-		rrs.rows = newRawRows()
-	}
-	if len(rrs.rows) == 0 {
-		rrs.updateFlushDeadline()
-	}
-	n := copy(rrs.rows[len(rrs.rows):cap(rrs.rows)], rows)
-	rrs.rows = rrs.rows[:len(rrs.rows)+n]
-	rows = rows[n:]
-	if len(rows) > 0 {
-		rowsToFlush = rrs.rows
-		rrs.rows = newRawRows()
-		rrs.updateFlushDeadline()
-		n = copy(rrs.rows[:cap(rrs.rows)], rows)
-		rrs.rows = rrs.rows[:n]
-		rows = rows[n:]
-	}
-	rrs.mu.Unlock()
-
-	return rows, rowsToFlush
-}
-
-func newRawRows() []rawRow {
-	return make([]rawRow, 0, maxRawRowsPerShard)
-}
-
 func (pt *partition) flushRowssToInmemoryParts(rowss [][]rawRow) {
-	if len(rowss) == 0 {
+	var nonEmptyRowss [][]rawRow
+	for _, rows := range rowss {
+		if len(rows) > 0 {
+			nonEmptyRowss = append(nonEmptyRowss, rows)
+		}
+	}
+	if len(nonEmptyRowss) == 0 {
 		return
 	}
 
 	// Convert rowss into in-memory parts.
-	var pwsLock sync.Mutex
-	pws := make([]*partWrapper, 0, len(rowss))
+	pws := make([]*partWrapper, len(nonEmptyRowss))
 	wg := getWaitGroup()
-	for _, rows := range rowss {
+	for i, rows := range nonEmptyRowss {
 		inmemoryPartsConcurrencyCh <- struct{}{}
-
 		wg.Go(func() {
-			pw := pt.createInmemoryPart(rows)
-			if pw != nil {
-				pwsLock.Lock()
-				pws = append(pws, pw)
-				pwsLock.Unlock()
-			}
-
+			pws[i] = pt.mustCreateInmemoryPart(rows)
 			<-inmemoryPartsConcurrencyCh
 		})
 	}
@@ -878,9 +742,14 @@ func (pt *partition) mustMergeInmemoryPartsFinal(pws []*partWrapper) *partWrappe
 	return newPartWrapperFromInmemoryPart(mpDst, flushToDiskDeadline)
 }
 
-func (pt *partition) createInmemoryPart(rows []rawRow) *partWrapper {
+// mustCreateInmemoryPart creates a new in-memory part from rawRows.
+//
+// The number of rawRows cannot be zero. Otherwise the method will panic.
+//
+// The returned value is always a non-nil partWrapper.
+func (pt *partition) mustCreateInmemoryPart(rows []rawRow) *partWrapper {
 	if len(rows) == 0 {
-		return nil
+		logger.Panicf("BUG: a part cannot be created from 0 rawRows")
 	}
 	mp := getInmemoryPart()
 	mp.InitFromRows(rows)
@@ -1142,7 +1011,7 @@ func (pt *partition) pendingRowsFlusher() {
 }
 
 func (pt *partition) flushPendingRows(isFinal bool) {
-	pt.rawRows.flush(pt, isFinal)
+	pt.rawRows.flush(pt.flushRowssToInmemoryParts, isFinal)
 }
 
 func (pt *partition) flushInmemoryRowsToFiles() {
@@ -1168,66 +1037,6 @@ func (pt *partition) flushInmemoryPartsToFiles(isFinal bool) {
 	}
 }
 
-func (rrss *rawRowsShards) flush(pt *partition, isFinal bool) {
-	var dst [][]rawRow
-
-	currentTimeMs := time.Now().UnixMilli()
-	flushDeadlineMs := rrss.flushDeadlineMs.Load()
-	if isFinal || currentTimeMs >= flushDeadlineMs {
-		rrss.rowssToFlushLock.Lock()
-		dst = rrss.rowssToFlush
-		rrss.rowssToFlush = nil
-		rrss.rowssToFlushLock.Unlock()
-	}
-
-	for i := range rrss.shards {
-		dst = rrss.shards[i].appendRawRowsToFlush(dst, currentTimeMs, isFinal)
-	}
-
-	pt.flushRowssToInmemoryParts(dst)
-}
-
-func (rrs *rawRowsShard) appendRawRowsToFlush(dst [][]rawRow, currentTimeMs int64, isFinal bool) [][]rawRow {
-	flushDeadlineMs := rrs.flushDeadlineMs.Load()
-	if !isFinal && currentTimeMs < flushDeadlineMs {
-		// Fast path - nothing to flush
-		return dst
-	}
-
-	// Slow path - move rrs.rows to dst.
-	rrs.mu.Lock()
-	dst = appendRawRowss(dst, rrs.rows)
-	rrs.rows = rrs.rows[:0]
-	rrs.mu.Unlock()
-
-	return dst
-}
-
-func (rrs *rawRowsShard) updateFlushDeadline() {
-	rrs.flushDeadlineMs.Store(time.Now().Add(pendingRowsFlushInterval).UnixMilli())
-}
-
-func appendRawRowss(dst [][]rawRow, src []rawRow) [][]rawRow {
-	if len(src) == 0 {
-		return dst
-	}
-	if len(dst) == 0 {
-		dst = append(dst, newRawRows())
-	}
-	prows := &dst[len(dst)-1]
-	n := copy((*prows)[len(*prows):cap(*prows)], src)
-	*prows = (*prows)[:len(*prows)+n]
-	src = src[n:]
-	for len(src) > 0 {
-		rows := newRawRows()
-		n := copy(rows[:cap(rows)], src)
-		rows = rows[:len(rows)+n]
-		src = src[n:]
-		dst = append(dst, rows)
-	}
-	return dst
-}
-
 func (pt *partition) mergePartsToFiles(pws []*partWrapper, stopCh <-chan struct{}, concurrencyCh chan struct{}, useSparseCache bool) error {
 	pwsLen := len(pws)
 

lib/storage/raw_row.go

@@ -3,11 +3,29 @@ package storage
 import (
 	"sort"
 	"sync"
+	"sync/atomic"
+	"time"
+	"unsafe"
 
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/atomicutil"
+	"github.com/VictoriaMetrics/VictoriaMetrics/lib/cgroup"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/decimal"
 	"github.com/VictoriaMetrics/VictoriaMetrics/lib/logger"
 )
 
+// The number of shards for rawRow entries.
+//
+// Higher number of shards reduces CPU contention and increases the max bandwidth on multi-core systems.
+var numRawRowsShards = cgroup.AvailableCPUs()
+
+// The interval for flushing buffered rows into parts, so they become visible to search.
+const pendingRowsFlushInterval = 2 * time.Second
+
+// The maximum number of rawRow items in rawRowsShard.
+//
+// Limit the maximum shard size to 8Mb, since this gives the lowest CPU usage under high ingestion rate.
+const maxRawRowsPerShard = (8 << 20) / int(unsafe.Sizeof(rawRow{}))
+
 // rawRow represents raw timeseries row.
 type rawRow struct {
 	// TSID is time series id.
@@ -149,3 +167,182 @@ func putRawRowsMarshaler(rrm *rawRowsMarshaler) {
 }
 
 var rrmPool sync.Pool
+
+type rawRowsShards struct {
+	flushDeadlineMs atomic.Int64
+
+	shardIdx atomic.Uint32
+
+	// Shards reduce lock contention when adding rows on multi-CPU systems.
+	shards []rawRowsShard
+
+	rowssToFlushLock sync.Mutex
+	rowssToFlush     [][]rawRow
+}
+
+func (rrss *rawRowsShards) init() {
+	rrss.shards = make([]rawRowsShard, numRawRowsShards)
+}
+
+func (rrss *rawRowsShards) Len() int {
+	n := 0
+	for i := range rrss.shards[:] {
+		n += rrss.shards[i].Len()
+	}
+
+	rrss.rowssToFlushLock.Lock()
+	for _, rows := range rrss.rowssToFlush {
+		n += len(rows)
+	}
+	rrss.rowssToFlushLock.Unlock()
+
+	return n
+}
+
+func (rrss *rawRowsShards) addRows(flush func([][]rawRow), rows []rawRow) {
+	shards := rrss.shards
+	shardsLen := uint32(len(shards))
+	for len(rows) > 0 {
+		n := rrss.shardIdx.Add(1)
+		idx := n % shardsLen
+		tailRows, rowsToFlush := shards[idx].addRows(rows)
+		rrss.addRowsToFlush(flush, rowsToFlush)
+		rows = tailRows
+	}
+}
+
+func (rrss *rawRowsShards) addRowsToFlush(flush func([][]rawRow), rowsToFlush []rawRow) {
+	if len(rowsToFlush) == 0 {
+		return
+	}
+
+	var rowssToMerge [][]rawRow
+
+	rrss.rowssToFlushLock.Lock()
+	if len(rrss.rowssToFlush) == 0 {
+		rrss.updateFlushDeadline()
+	}
+	rrss.rowssToFlush = append(rrss.rowssToFlush, rowsToFlush)
+	if len(rrss.rowssToFlush) >= defaultPartsToMerge {
+		rowssToMerge = rrss.rowssToFlush
+		rrss.rowssToFlush = nil
+	}
+	rrss.rowssToFlushLock.Unlock()
+
+	flush(rowssToMerge)
+}
+
+func (rrss *rawRowsShards) updateFlushDeadline() {
+	rrss.flushDeadlineMs.Store(time.Now().Add(pendingRowsFlushInterval).UnixMilli())
+}
+
+func (rrss *rawRowsShards) flush(flush func(rrs [][]rawRow), isFinal bool) {
+	var dst [][]rawRow
+
+	currentTimeMs := time.Now().UnixMilli()
+	flushDeadlineMs := rrss.flushDeadlineMs.Load()
+	if isFinal || currentTimeMs >= flushDeadlineMs {
+		rrss.rowssToFlushLock.Lock()
+		dst = rrss.rowssToFlush
+		rrss.rowssToFlush = nil
+		rrss.rowssToFlushLock.Unlock()
+	}
+
+	for i := range rrss.shards {
+		dst = rrss.shards[i].appendRawRowsToFlush(dst, currentTimeMs, isFinal)
+	}
+
+	flush(dst)
+}
+
+type rawRowsShardNopad struct {
+	flushDeadlineMs atomic.Int64
+
+	mu   sync.Mutex
+	rows []rawRow
+}
+
+type rawRowsShard struct {
+	rawRowsShardNopad
+
+	// The padding prevents false sharing
+	_ [atomicutil.CacheLineSize - unsafe.Sizeof(rawRowsShardNopad{})%atomicutil.CacheLineSize]byte
+}
+
+func (rrs *rawRowsShard) Len() int {
+	rrs.mu.Lock()
+	n := len(rrs.rows)
+	rrs.mu.Unlock()
+	return n
+}
+
+func (rrs *rawRowsShard) addRows(rows []rawRow) ([]rawRow, []rawRow) {
+	var rowsToFlush []rawRow
+
+	rrs.mu.Lock()
+	if cap(rrs.rows) == 0 {
+		rrs.rows = newRawRows()
+	}
+	if len(rrs.rows) == 0 {
+		rrs.updateFlushDeadline()
+	}
+	n := copy(rrs.rows[len(rrs.rows):cap(rrs.rows)], rows)
+	rrs.rows = rrs.rows[:len(rrs.rows)+n]
+	rows = rows[n:]
+	if len(rows) > 0 {
+		rowsToFlush = rrs.rows
+		rrs.rows = newRawRows()
+		rrs.updateFlushDeadline()
+		n = copy(rrs.rows[:cap(rrs.rows)], rows)
+		rrs.rows = rrs.rows[:n]
+		rows = rows[n:]
+	}
+	rrs.mu.Unlock()
+
+	return rows, rowsToFlush
+}
+
+func newRawRows() []rawRow {
+	return make([]rawRow, 0, maxRawRowsPerShard)
+}
+
+func (rrs *rawRowsShard) updateFlushDeadline() {
+	rrs.flushDeadlineMs.Store(time.Now().Add(pendingRowsFlushInterval).UnixMilli())
+}
+
+func (rrs *rawRowsShard) appendRawRowsToFlush(dst [][]rawRow, currentTimeMs int64, isFinal bool) [][]rawRow {
+	flushDeadlineMs := rrs.flushDeadlineMs.Load()
+	if !isFinal && currentTimeMs < flushDeadlineMs {
+		// Fast path - nothing to flush
+		return dst
+	}
+
+	// Slow path - move rrs.rows to dst.
+	rrs.mu.Lock()
+	dst = appendRawRowss(dst, rrs.rows)
+	rrs.rows = rrs.rows[:0]
+	rrs.mu.Unlock()
+
+	return dst
+}
+
+func appendRawRowss(dst [][]rawRow, src []rawRow) [][]rawRow {
+	if len(src) == 0 {
+		return dst
+	}
+	if len(dst) == 0 {
+		dst = append(dst, newRawRows())
+	}
+	prows := &dst[len(dst)-1]
+	n := copy((*prows)[len(*prows):cap(*prows)], src)
+	*prows = (*prows)[:len(*prows)+n]
+	src = src[n:]
+	for len(src) > 0 {
+		rows := newRawRows()
+		n := copy(rows[:cap(rows)], src)
+		rows = rows[:len(rows)+n]
+		src = src[n:]
+		dst = append(dst, rows)
+	}
+	return dst
+}