标签：Google target zone Database time range Time Monarch data

Google 20年发的时序库的paper

首先看看系统架构，

设计时，首先考虑CAP的balance问题，向Spanner这样的高一致性方案，明显不适合监控系统，所以Monarch牺牲一致性来保证实时性和可用性

Monarch’s design is determined by its primary usage for monitoring and alerting.
First, Monarch readily trades consistency for high availability and partition tolerance [21, 8, 9].
Writing to or reading from a strongly consistent database like Spanner [13] may block for a long time;
that is unacceptable for Monarch because it would increase mean-time-to-detection and mean-time-to-mitigation for potential outages.
To promptly deliver alerts, Monarch must serve the most recent data in a timely fashion;
for that, Monarch drops delayed writes and returns partial data for queries if necessary.
In the face of network partitions, Monarch continues to support its users’ monitoring and alerting needs, with mechanisms to indicate the underlying data may be incomplete or inconsistent.

第二，Monarch降低对于其他google系统的依赖，因为这样回循环依赖，所以选择将数据存在内存里面。

Second, Monarch must be low dependency on the alerting critical path.
To minimize dependencies, Monarch stores monitoring data in memory despite the high cost.
Most of Google’s storage systems, including Bigtable [10], Colossus ([36], the successor to GFS [20]), Spanner [13], Blobstore [18], and F1 [40], rely on Monarch for reliable monitoring;
thus, Monarch cannot use them on the alerting path to avoid a potentially dangerous circular dependency.
As a result, non-monitoring applications (e.g., quota services) using Monarch as a global time series database are forced to accept reduced consistency. 

首先Highlevel上，是分层的，从Global分到各个zone，这个很容易理解

Monarch的架构，分成三种颜色，

分别代表，

persistence层，

- Leave，内存时序库；

- Recovery logs，WAL？可以回放出data的log，可以用于长期persistence保存

- 配置服务，用Spanner来存储配置数据

Ingestion层，

- Ingestion routers和Leaf routers，两层路由

- Range assigners，管理range分布的，在leaf间做负载均衡

Query层，

- Mixer，reducer，query broker，会合并所有子查询的结果

- Index Server，索引，每个leaf里面哪些数据，用于skip的

- Evaluator，做物化视图的，定期执行特定查询写回leave

 

再者，看下数据模型，存储格式

首先Monarch是schematized tables 

并且分成两部分，Target Schema和Metric Schema，类似Timon

Target，如其名，用来记录这条消息的source，这里图里的例子是ComputeTask

类似timon的tag，这里Target是重复记录的，有点低效

Monarch uses targets to associate each time series with its source entity (or monitored entity), which is, for example, the process or the VM that generates the time series.
Each target represents a monitored entity, and conforms to a target schema that defines an ordered set of target field names and associated field types.
Figure 2 shows a popular target schema named ComputeTask;
each ComputeTask target identifies a running task in a Borg [46] cluster with four fields: user, job, cluster, and task num.

为了locality，需要指定一个字段做location

For locality, Monarch stores data close to where the data is generated.
Each target schema has one field annotated as location;
the value of this location field determines the specific Monarch zone to which a time series is routed and stored.
For example, the location field of ComputeTask is cluster; each Borg cluster is mapped to one (usually the closest) Monarch zone.
As described in Section 5.3, location fields are also used to optimize query execution.

Within each zone, Monarch stores time series of the same target together in the same leaf because they originate from the same entity and are more likely to be queried together in a join.
Monarch also groups targets into disjoint target ranges in the form of [Sstart,Send) where Sstart and Send are the start and end target strings.
A target string represents a target by concatenating the target schema name and field values in order1.
For example, in Figure 2, the target string ComputeTask::sql-dba::db.server::aa::0876 represents the Borg task of a database server.
Target ranges are used for lexicographic sharding and load balancing among leaves (see Section 4.2);
this allows more efficient aggregation across adjacent targets in queries (see Section 5.3). 

Metrics就很容易理解，就是时序点，这里的例子不太好，他给出的是一个特殊类型，Distribution，就是histogram，所以是个map；正常就是一个数值

A metric measures one aspect of a monitored target, such as the number of RPCs a task has served, the memory utilization of a VM, etc.
Similar to a target, a metric conforms to a metric schema, which defines the time series value type and a set of metric fields.
Metrics are named like files.
Figure 2 shows an example metric called /rpc/server/latency that measures the latency of RPCs to a server; it has two metric fields that distinguish RPCs by service and command.

The value type can be boolean, int64, double, string, distribution, or tuple of other types.
All of them are standard types except distribution, which is a compact type that represents a large number of double values.
A distribution includes a histogram that partitions a set of double values into subsets called buckets and summarizes values in each bucket using overall statistics such as mean, count, and standard deviation [28]. Bucket boundaries are configurable for trade-off between data granularity (i.e., accuracy) and storage costs: users may specify finer buckets for more popular value ranges.
Figure 3 shows an example distribution-typed time series of /rpc/server/latency which measures servers’ latency in handling RPCs;
and it has a fixed bucket size of 10ms.

 

接着看写入路径，ingestion，

首先两级partition，一级是根据location，一级是根据range assigner

写入过程没啥特殊的，内存存储做了一些优化，主要是timestamp sharing和delta encoding

The right side of Figure 1 gives an overview of Monarch’s collection path.
The two levels of routers perform two levels of divide-and-conquer:
ingestion routers regionalize time series data into zones according to location fields,
and leaf routers distribute data across leaves according to the range assigner.
Recall that each time series is associated with a target and one of the target fields is a location field. 

 

 

Range assigner，负载均衡，在zone内会根据cpu和memory，去split，merge，move range

这里有点意思的是，异构的副本策略，副本数目和时间粗细粒度是可以不同的

In addition, we allow heterogeneous replication policies (1 to 3 replicas) for users to trade off between availability and storage cost.
Replicas of each target range have the same boundaries, but their data size and induced CPU load may differ because,
for example, one user may retain only the first replica at a fine time granularity while another user retains all three replicas at a coarse granularity.
Therefore, the range assigner assigns each target range replica individually.
Of course, a leaf is never assigned multiple replicas of a single range.
Usually, a Monarch zone contains leaves in multiple failure domains (clusters); the assigner assigns the replicas for a range to different failure domains. 

Range assigners balance load in ways similar to Slicer [1].
Within each zone, the range assigner splits, merges, and moves ranges between leaves to cope with changes in the CPU load and memory usage imposed by the range on the leaf that stores it.
While range assignment is changing, data collection works seamlessly by taking advantage of recovery logs. 

 

Aggregation，聚合

核心的一点是，对于监控场景，时序数据量太大了，而且没有必要记录原始值，所以在ingestion的时候直接聚合就好，和Timon思路一致

For some monitoring scenarios, it is prohibitively expensive to store time series data exactly as written by clients.
One example is monitoring disk I/O, served by millions of disk servers, where each I/O operation (IOP) is accounted to one of tens of thousands of users in Google.
This generates tens of billions of time series, which is very expensive to store naively.
However, one may only care about the aggregate IOPs per user across all disk servers in a cluster.
Collection aggregation solves this problem by aggregating data during ingestion. 

 

聚合的思路，大同小异，他这里没有提Dataflow， 

内存中维持滑动窗口，窗口解决乱序，触发后就变成immutable，late数据直接丢掉，聚合完的数据写入内存库和recovery logs里面。

这里有个问题，内存窗口里面的数据会丢啊，多个副本保证吗？

In addition, each leaf also maintains a sliding admission window and rejects deltas older than the window length TW .
Therefore, older buckets become immutable and generate finalized points that can be efficiently stored with delta and run-length encoding.
The admission window also enables Monarch to recover quickly from network congestion;
otherwise, leaves may be flooded by delayed traffic and never catch up to recent data, which is more important for critical alerting.
In practice, rejected writes comprise only a negligible fraction of traffic.
Once a bucket’s end time moves out of the admission window, the bucket is finalized: the aggregated point is written to the in-memory store and the recovery logs. 

 

 

最后看查询层，

除了普通的查询，

还支持Standing查询，其实就是定期下发的查询，有两种左右：

物化视图的生成，这个挺有意思，这样生成物化视图比较简单，但是一致性保证不了

alerts，实时性会有问题

Standing queries are periodic materialized-view queries whose results are stored back into Monarch; teams use them:
(1) to condense data for faster subsequent querying and/or cost-saving; and (2) to generate alerts.
Standing queries can be evaluated by either regional zone evaluators or global root evaluators.
The decision is based on static analysis of the query and the table schemas of the inputs to the query (details in Section 5.3).
The majority of standing queries are evaluated by zone evaluators which send identical copies of the query to the corresponding zone mixers and write the output to their zone.
Such queries are efficient and resilient to network partition.
The zone and root evaluators are sharded by hashes of standing queries they process, allowing us to scale to millions of standing queries. 

 

后面还讲了些，分层查询，pushdown，索引，配置管理，不一一写了

 

标签：Google,target,zone,Database,time,range,Time,Monarch,data
来源： https://www.cnblogs.com/fxjwind/p/16166508.html

本站声明： 1. iCode9 技术分享网（下文简称本站）提供的所有内容，仅供技术学习、探讨和分享；
2. 关于本站的所有留言、评论、转载及引用，纯属内容发起人的个人观点，与本站观点和立场无关；
3. 关于本站的所有言论和文字，纯属内容发起人的个人观点，与本站观点和立场无关；
4. 本站文章均是网友提供，不完全保证技术分享内容的完整性、准确性、时效性、风险性和版权归属；如您发现该文章侵犯了您的权益，可联系我们第一时间进行删除；
5. 本站为非盈利性的个人网站，所有内容不会用来进行牟利，也不会利用任何形式的广告来间接获益，纯粹是为了广大技术爱好者提供技术内容和技术思想的分享性交流网站。