今天，總部位于波士頓的 Metaplane 是一家致力于改善和糾正企業(yè)數(shù)據(jù)質(zhì)量問題的初創(chuàng)公司，宣布已在 A 輪融資中籌集了 1380 萬美元。

風(fēng)險(xiǎn)投資公司 Felicis 領(lǐng)投，Khosla Ventures、Flybridge、Y Combinator、Stage 2 Capital、B37 和 SNR 跟投。

Metaplane 表示，計(jì)劃利用本輪融資進(jìn)一步開發(fā)其人工智能驅(qū)動(dòng)的數(shù)據(jù)可觀測(cè)平臺(tái)，并成為“無可爭(zhēng)議的最強(qiáng)大、可配置和神奇的信任數(shù)據(jù)解決方案”。

該公司由麻省理工學(xué)院畢業(yè)生 Kevin Hu、前 HubSpot 工程師 Peter Casinelli 和前 Appcues 開發(fā)人員 Guru Mahendran 創(chuàng)立，正在快速發(fā)展的數(shù)據(jù)可觀測(cè)性領(lǐng)域與 Monte Carlo、Observe 和 Acceldata 等資金雄厚的參與者展開競(jìng)爭(zhēng)。

去年，該公司的客戶群增長(zhǎng)了三倍，并已與 Bose、Sigma、Klaviyo 和 ClickUp 等品牌合作。

監(jiān)控和標(biāo)記整個(gè)數(shù)據(jù)堆棧的問題

數(shù)據(jù)已成為現(xiàn)代企業(yè)的驅(qū)動(dòng)力，使團(tuán)隊(duì)不僅能夠分析決策的歷史模式，還能預(yù)測(cè)增長(zhǎng)的關(guān)鍵方面，例如特定事件的庫(kù)存計(jì)劃。

生成式人工智能應(yīng)用程序的激增也促使公司將不同來源的數(shù)據(jù)整合在一起，并有望推動(dòng)進(jìn)一步的價(jià)值。

然而，鑒于這種向數(shù)據(jù)驅(qū)動(dòng)工作的巨大轉(zhuǎn)變，團(tuán)隊(duì)很難密切關(guān)注他們所掌握的有關(guān)質(zhì)量問題的所有信息。

管道變得更加復(fù)雜，有時(shí)需要處理數(shù)百或數(shù)千個(gè)來源。

Metaplane 將人工智能應(yīng)用于這個(gè)問題，據(jù)稱這使企業(yè)能夠主動(dòng)監(jiān)視其數(shù)據(jù)生態(tài)系統(tǒng)不同層的數(shù)據(jù)事件。

“我們集成了盡可能多的數(shù)據(jù)堆棧，無論是 Fivetran 等攝取工具、Snowflake 和 BigQuery 等云數(shù)據(jù)倉(cāng)庫(kù)、dbt 和 Airflow 等轉(zhuǎn)換和編排層、Census 和 Hightouch 等反向 ETL 工具，以及 Sigma 等 BI 工具、Tableau 和 Looker。

我們更進(jìn)一步，成為唯一與 Postgres 和 MySQL 等事務(wù)數(shù)據(jù)庫(kù)集成的數(shù)據(jù)可觀察性產(chǎn)品，并捕獲 Github 中 dbt Pull 請(qǐng)求中的問題?！盚u 于 2019 年從 MIT 的一個(gè)項(xiàng)目創(chuàng)辦了這家公司，他告訴 VentureBeat。

通過機(jī)器學(xué)習(xí)監(jiān)控?cái)?shù)據(jù)質(zhì)量

一旦平臺(tái)與數(shù)據(jù)堆棧集成，用戶就可以在頻繁使用/更新的表上設(shè)置監(jiān)視器，以密切關(guān)注不同的數(shù)據(jù)質(zhì)量指標(biāo)，例如新鮮度、行數(shù)、唯一性和空值。

整個(gè)過程大約需要 15 分鐘，隨后產(chǎn)品開始與 AI 配合使用。

As Hu explained, the system’s machine learning (ML) model trains on the data profile, using historical metadata, and then starts flagging data anomalies (even schema changes) within a day or two. The whole thing is fully automated, with alerts going directly to concerned data teams on the preferred destination for alerts.

“We use the most historical data to train our models, ensuring that we can capture seasonality and avoid repetitive alerts. Every business is unique and simply applying a one-size-fits-all model to each customer introduces a lot of inaccuracy. Unlike other monitoring tools, we also make it easy for users to tweak models to ignore one-offs or learn new trends to account for seasonal patterns and factors specific to their industry. Customers go with us because we catch issues that others can’t while keeping the noise to a minimum,” Hu explained.

Notably, in addition to monitoring metrics like freshness and volume of data, Metaplane can also go deeper to detect data problems that are very domain-specific with finer-grain controls, including monitoring for changes in data usage and cloud warehouse spend. Plus, the coverage of the data stack allows the platform to create a complete picture of column-level lineage from data source to destination and provide context on the downstream impact of issues as well as upstream root causes.

80,000 data quality incidents resolved

While Metaplane is not as heavily funded as its competitors Observe, Acceldata and Monte Carlo, the company has been doing pretty well in the data observability space. In 2023, its ARR grew six-fold while the customer base grew three-fold to over 100 enterprises – with known names like Klaviyo, Bose, ClickUp, Sigma, Census, GoFundMe and Ramp coming on board.

As of January 2024, the company said, these customers had run 500 million data quality checks on over 40 million data assets and over 30 million data lineage connections, detecting and resolving as many as 80,000 incidents.

“We believe that all companies should be able to trust their data, and so we enable teams to sign up and use it for free. As a result, we’ve benefited greatly from organic growth and more users have used Metaplane than any other data observability tool,” the founder emphasized.

In addition to the self-serve approach to adoption, Hu claimed that the platform’s ability to detect important issues while keeping noise to the minimum and give a complete view of the data stack makes it better than all other observability tools out there.

“我是否監(jiān)控了所有可能給數(shù)據(jù)帶來錯(cuò)誤的事情？

有多少問題源于事務(wù)數(shù)據(jù)庫(kù)？

有多少可以通過阻止代碼更改來阻止？

回答這些問題的唯一方法是在整個(gè)數(shù)據(jù)堆棧中，在所有可能產(chǎn)生或影響數(shù)據(jù)問題的地方進(jìn)行深度集成。

我們最近宣布與兩個(gè)領(lǐng)先的反向 ETL 平臺(tái) Census 和 Hightouch 進(jìn)??行集成，并且很快還會(huì)發(fā)布更多公告?！盚u 補(bǔ)充道。

展望未來，該公司計(jì)劃利用這筆資金專注于研發(fā)，并進(jìn)一步開發(fā)其數(shù)據(jù)可觀測(cè)平臺(tái)，讓企業(yè)團(tuán)隊(duì)能夠放心地使用其數(shù)據(jù)資產(chǎn)。

其中一部分將致力于實(shí)現(xiàn)更多監(jiān)控架構(gòu)的自動(dòng)化，同時(shí)引入對(duì)觀察更多指標(biāo)、來源和來源之間連接的支持。

“我們的愿景是，我們的平臺(tái)將了解每個(gè)客戶的獨(dú)特需求，并根據(jù)他們不斷變化的需求推薦理想的監(jiān)控和警報(bào)架構(gòu)。

我們將把這一點(diǎn)與我們監(jiān)控的內(nèi)容的廣泛擴(kuò)展相結(jié)合，添加更深入的指標(biāo)和更廣泛的指標(biāo)，以觀察數(shù)據(jù)堆棧中的所有內(nèi)容，以便我們的客戶始終擁有必要的上下文來查找和解決數(shù)據(jù)質(zhì)量問題，”Hu 指出。