Techbanq

Platforms · search and analytics

Search that cannot read Arabic is not finished in this market

A search deployment succeeds on three fronts at once: relevance a user in Riyadh recognises in either language, an index lifecycle that keeps last year’s logs from billing like this morning’s, and a cluster sized from measured workloads rather than a formula found online. We engineer all three, and we will also tell you when your database’s own search is enough.

Serving Saudi Arabia from our Riyadh office and our engineering hubs in Lahore and Karachi

Market
Saudi Arabia
Written for
CTO, Head of Digital
Decision
Build search and log analytics that stay relevant, affordable and secure
Languages
EN · AR

Enterprise delivery credentials

  • ISO

    ISO/IEC 27001

    ISO/IEC 27001 certified

Certifications & registrations

The economics

What actually decides the cost of running Elasticsearch

Search clusters do not get expensive at purchase; they get expensive through defaults nobody revisited. These are the levers, and every one of them is a policy decision wearing a technical costume.

VariableEffect on priceWhy
Retention without a lifecycleThe largest bill in most estatesLogs indexed once and never expired accumulate until storage, not traffic, is what you are paying for. An index lifecycle policy — hot to warm to cold to deleted, on dates the business agreed — is the single highest-return piece of configuration in the platform.
Replicas chosen by habitStorage and memory multiplied silentlyEvery replica doubles what an index consumes. Production search that must survive a node loss earns its replicas; a development index or a re-buildable analytics index often does not, and paying production redundancy for throwaway data is pure waste.
Mappings that index everythingLarger indices, slower ingest, no benefitBy default every field is made searchable, including the many nobody will ever query. Mapping only what is searched and storing the rest shrinks the index and speeds both writes and queries — a design pass that pays for itself in the first month.
Queries written against the grainA cluster sized for its worst queryDeep pagination, unbounded aggregations and wildcard-heavy queries force the cluster to do work its design punishes. Fixing the query patterns is routinely cheaper than the node the dashboard says you need.

Ask any supplier quoting this

  • Which of our indices have no lifecycle policy, and what are they costing per month?
  • What share of our storage is replicas, and which of them protect data we could rebuild?
  • Which ten queries consume the most cluster time, and who owns each?

Relevance

Arabic analysis is relevance engineering, not localisation

Full-text search works by analysis: text is broken into tokens, normalised and reduced toward root forms, so that a query and a document can meet even when their surface forms differ. Arabic makes every step of that pipeline do real work. The same word arrives with or without hamza variants, with letters that writers use interchangeably, with diacritics present or absent, and inflected by prefixes and suffixes that attach articles, prepositions and pronouns to the stem. An index built with the default analyzer treats each variant as a different word, and the user experience follows: a search that finds a document only when the spelling matches exactly is, to its users, a search that does not work.

The engineering answer exists and is well understood: Arabic-aware analysis that normalises the variant characters, strips diacritics, and stems the attached forms back to something queries can meet — tuned, because over-aggressive stemming collapses words that should stay distinct. What does not work is bolting it on later: analysis is decided at index time, so getting Arabic right after launch means reindexing the estate. That is why we treat Arabic analysis as a first-class design input, tested against real bilingual queries from day one, and measured the only way relevance can be — with judged result sets, before and after.

Sizing

Sizing a cluster from measurements instead of folklore

Every sizing formula circulating online encodes someone else’s workload. Yours is knowable — by measuring it — and a cluster sized from measurement is the difference between paying for headroom and paying for guesswork.

STAGE 01

Characterise the workload you actually have

Ingest rate and its peaks, document sizes, query shapes and their latency expectations, retention per data stream — captured from real systems, because each of these moves the answer more than any tuning parameter. A workload description that is data, not adjectives.

STAGE 02

Replay it against a candidate shape

A test cluster receives representative data and the real query mix at production rates, pushed until it degrades — because the number you need is where it breaks, not whether it works at yesterday’s volume. A measured ceiling per node type and shard layout.

STAGE 03

Size from the measurement, with declared headroom

The production cluster is the measured requirement plus growth you can defend and failure capacity you have chosen — each stated, so the cost of the margin is a decision someone made rather than padding nobody owns. A cluster whose every node has a reason on paper.

STAGE 04

Re-measure when the workload moves

A new data stream, a new dashboard, a doubling of users — each invalidates the old measurement. The replay harness stays, so resizing is an afternoon of evidence rather than a season of debate. Capacity that tracks reality instead of trailing it.

Securing it

The controls a search cluster is not production without

Search clusters concentrate an organisation’s most searchable data in one place, and unsecured clusters reachable from the internet remain one of the most common ways large datasets leak. None of these controls is exotic; all of them have to be on.

No network path from the internet

The classic leak is not a clever exploit; it is a cluster that answered on a public address to anyone who asked. The cluster lives on private networks, reached through the applications that front it.

The state it leaves you in: A dataset that cannot be downloaded by a scanner.

Authentication and encryption everywhere

Every request authenticated, transport encrypted between nodes and clients alike — the platform ships these; leaving them off is a choice, and the wrong one.

The state it leaves you in: No anonymous reads, no plaintext data in flight.

Authorisation down to indices and fields

Role-based access scoped per index — and per field where a document mixes sensitivities — so the analytics team reads analytics, not the HR index that shares the cluster.

The state it leaves you in: Blast radius bounded by role, not by cluster.

An audit trail on the data plane

Who queried what, and when, recorded — the question an investigation asks first, and unanswerable unless it was switched on before the incident.

The state it leaves you in: An incident you can reconstruct instead of estimate.

Our role

What we engineer on the search platform and what we tell you not to build

  1. 01

    The honest threshold

    PostgreSQL full-text search or your database’s own capabilities cover more use cases than the search industry admits. Where your need is a filtered lookup over modest data, we say so and save you a cluster; a dedicated search platform is recommended when relevance ranking, language analysis or log volume genuinely demand it.

  2. 02

    Relevance measured in both languages

    We design the analysis chain for Arabic and English together, and prove it against judged sets of real user queries — because relevance claimed without measurement is opinion with a dashboard.

  3. 03

    A platform your operators can afford

    Lifecycle policies, mapping discipline, the sizing harness and the security baseline are delivered as configuration and runbooks your team owns, so the cluster’s costs and controls remain legible after we leave.

What your users should find first, and how long each class of data must be kept, are product and policy calls that stay with you; we build the analysis, lifecycle and sizing that execute those calls at a cost you chose on purpose.

Common questions

What buyers ask

Is our database’s built-in search enough, or do we need Elasticsearch?+

Start from what the search must do, not from what the tools can. Database full-text search — PostgreSQL’s in particular — handles exact and prefix matching, basic ranking and filtered lookups well, with zero extra infrastructure, and for an internal tool or a catalogue of thousands of records it is often the right and sufficient answer. A dedicated search platform earns its operational cost when you need relevance ranking users will judge you by, language analysis such as proper Arabic stemming, typo tolerance, faceting at scale, or log analytics over volumes a transactional database should not carry. The expensive mistake in both directions is the same one: adopting a cluster for a lookup, or torturing a database into being a search engine.

Why does our search work badly in Arabic when it works fine in English?+

Almost certainly because the index was built with an analyzer designed for English-like text. Arabic needs its own analysis: normalising the characters writers use interchangeably, stripping diacritics, and stemming the prefixes and suffixes that attach articles, prepositions and pronouns to words. Without that, each surface form of the same word is indexed as a different term, so users only find documents whose spelling happens to match their query exactly — which reads, to them, as broken search. The fix is real but not a toggle: analysis is applied at index time, so the analyzer chain must be configured, tested against genuine Arabic queries, and the affected indices rebuilt. It is precisely why we treat Arabic analysis as part of initial design rather than a fix scheduled for later.

How do you size an Elasticsearch cluster without overpaying for it?+

By refusing to size it from a formula. Sizing rules of thumb encode somebody else’s document sizes, query shapes and retention, and applying them to your workload produces a number that is confidently wrong in an unknown direction. Our method is to measure: characterise the real workload from your systems, replay representative data and the actual query mix against a candidate cluster at production rates, and push it until it degrades so the ceiling is a measurement. Production capacity is then that measurement plus explicitly chosen headroom for growth and node failure — with each margin stated, so every riyal of the cluster is attached to a reason. And because workloads move, the replay harness stays with your team, making the next resize an exercise in evidence rather than a renegotiation from scratch.

Next step

Measure your search workload

A workload review that measures what you actually index and query, proposes the lifecycle and sizing that fit it, and tests bilingual relevance against real user searches.